Balancing accuracy and efficiency in lumbar spine image segmentation using multi-scale attention and residual learning Authors: Vijaya Bhaskar Sadu, Bharathi Subramaniam, Rukmani Devi Sethuraman, Sudhakar Tummala, Beluguri Venkateswarlu, Jeevika Tharini Viswanathan Published: 2026 ABSTRACT AbstractAccurate automated segmentation of Lumbar Spine Structures (LSS) in Magnetic Resonance Imaging (MRI) is important for effective diagnosis and treatment planning. However, most current methods encounter a trade-off between segmentation precision and computational efficiency. In particular, several methods fall short in modelling the intricate multi-scale anatomical details and subtle boundary transitions typical of vertebral bodies and intervertebral discs, thereby limiting their clinical utility. This study introduces Multiscale Residual and Adaptive Spatial-Channel Feature Pyramid Network (MRAS + FPN), a novel Deep Learning (DL) designed to enhance Lumbar Spine Segmentation. The model integrates Multi-Scale Attention (MS) with Residual Learning (RL) to balance accuracy and efficiency. Two principal innovations are embedded in the design: (a) Multiscale Residual Attention Blocks (MRAB), which incorporate Directionally Dilated Convolution Modules (DDCM) coupled with spatial-channel attention to address anisotropic anatomical structures, and (b) Adaptive Spatial-Channel Attention Blocks (ASCAB), which refine Feature Selection (FS) across pyramid levels dynamically. Also, Cross-level Feature Gating (CFG) is employed to regulate information flow between the encoder and decoder. The model was trained and validated on the MRSpineSeg dataset, which comprises 215 T2-weighted lumbar MRI volumes annotated by experts, covering vertebrae T9-S1 and their corresponding intervertebral discs. A composite loss function was used, combining Dice, Binary Cross-Entropy, and a boundary-sensitive term. MRAS-FPN proved to have improved segmentation performance compared with established models, including 2D U-Net, 3D U-Net, and DeepLab v3+. It achieved a Dice Similarity Coefficient (DSC) of 87.8 ± 1.3% and a mean Intersection over Union (mIoU) of 78.2 ± 2.1%, outperforming the best baseline (3D U-Net: 83.4% DSC, 73.4% mIoU) by 4.4% and 4.8%. Evaluation of boundary accuracy showed marked improvements, with a 95th percentile Hausdorff Distance (HD95) of 3.18 ± 0.8 mm and an Average Symmetric Surface Distance (ASSD) of 1.32 ± 0.5 mm—representing gains of 19.9% and 25.0% compared to 3D U-Net. Structure-wise analysis indicated stable segmentation quality across all vertebral levels (T9–S1: 84.4% mean DSC) and intervertebral discs (85.6% mean DSC), with notable gains in regions of high anatomical complexity. Ablation experiments further confirmed the utility of each component, with MRAB generating a 3.9% improvement in DSC and ASCAB contributing an additional 2.1% over the base FPN. MRAS-FPN achieves a well-calibrated balance between segmentation fidelity and resource efficiency, generating SOTA results while requiring manageable computational resources (22.1 hours of training time and 15.8 GB of memory). Its MSA successfully captures intricate anatomical features and boundary definitions critical for clinical use. The model’s strong performance across a standard range of spinal structures and its resilience to overfitting underscore its applicability in real-world scenarios, including surgical planning, diagnostic support, and automated testing of the Lumbar Spine (LS).Keywords: Magnetic resonance imaging, Deep learning, Lumbar spine structures, Residual learning, Multi-scale attention, Diagnostic support systemSubject terms: Anatomy, Computational biology and bioinformatics, Engineering, Health care, Mathematics and computing, Medical research Introduction Lumbar Spine (LS) disorders are among the most common musculoskeletal conditions globally, with estimates indicating that nearly 80% of individuals experience some form of lower back pain during their lifetime, resulting in significant healthcare costs and productivity losses1,2. Magnetic Resonance Imaging (MRI) has become indispensable for the comprehensive analysis of Lumbar Spine Structures (LSS), providing critical insights that inform diagnosis, therapeutic planning, and longitudinal monitoring of spinal pathologies such as disc herniation, spinal stenosis, degenerative disc disease, and vertebral compression fractures3–5. The ability to segment vertebral bodies and intervertebral discs accurately from MRI data is fundamental for enabling robust quantitative analysis, deriving volumetric indices, and powering computer-aided diagnostic tools designed to enhance Clinical Decision-Making Systems (CDMS)6–10. Despite being the current clinical standard, manual segmentation of LSS in MRI poses several challenges that constrain its broader utility and scalability11–13. The task is notably time-intensive, frequently requiring 2–3 h for experienced radiologists to manually delineate vertebrae and intervertebral discs across all sagittal slices of a typical scan14–16. This labor-intensive process creates substantial bottlenecks in busy radiology departments, where timely image interpretation is crucial to maintaining continuity of care17,18. Moreover, manual annotation is inherently subjective, leading to notable inter-observer and intra-observer variability; reported Dice coefficients among expert annotators typically range from 0.75 to 0.8511–13. Such variability compromises diagnostic reliability and impacts downstream treatment outcomes4,13. These limitations are even more pronounced in pathological cases, such as advanced disc degeneration or vertebral compression fractures, where distorted or occluded anatomical boundaries make reliable manual interpretation particularly difficult19. A range of automated segmentation methods has been developed to mitigate the limitations of manual annotation, yet many face persistent challenges that hinder their routine clinical deployment17,18,20,21. Classical image processing methods, such as edge detection, region growing, and atlas-based approaches, have been applied sparingly in LS-MRI, primarily due to the marked anatomical variability of the region and the subtle intensity gradients between adjacent structures22,23. These traditional approaches frequently report Dice scores below 75% and tend to fail under pathological conditions or in the presence of imaging objects, leading to unreliable segmentation outputs23. More recently, deep learning, particularly CNN-based architectures such as U-Net and its variants, has delivered notable gains in segmentation accuracy but still exhibits critical shortcomings17,18. The standard 2D U-Net frequently fails to incorporate sufficient inter-slice spatial context, leading to discontinuities across adjacent slices24. Although 3D U-Net improves spatial coherence, it remains computationally demanding and continues to struggle with fine boundary delineation24–26. Feature Pyramid Networks (FPN) and attention-enhanced models such as DeepLab v3 + are generally not tailored to capture the directional and anisotropic features unique to spinal anatomy14,18,27. The persistent gap between existing automated segmentation methods and the clinical demands of spinal imaging underscores the need for more principled solutions that balance high accuracy with computational feasibility. The intricate anatomy of the LS introduces distinct challenges that require carefully designed network architectures. Vertebral bodies and intervertebral discs exhibit marked morphological variability across spinal levels (T9-S1), and subtle intensity gradients in T2-weighted MRI scans require models with broad, multi-scale feature representation3,28. These structures also display marked anisotropy: vertebral bodies tend to exhibit a vertically aligned morphology, whereas intervertebral discs exhibit horizontal elongation, posing specific demands on FE that must be sensitive to directional patterns5,20. Pathological conditions such as disc degeneration or vertebral fractures frequently distort anatomical regularity, making it essential for segmentation models to incorporate adaptive attention mechanisms that generalise across a broad spectrum of clinical scenarios10,19. As the clinical demand for quantitative spine analysis continues to rise alongside an ongoing shortage of radiology specialists, there is a pressing requirement for robust automated segmentation tools that can deliver expert-level performance while remaining computationally feasible for integration into routine diagnostic workflows4,21. This study aims to develop a novel Deep Learning (DL) model that effectively balances segmentation accuracy with computational efficiency in automated LS-MRI analysis. The overarching goal is to construct a Multi-Scale Attention (MSA) capable of accurately delineating vertebral bodies and intervertebral discs while operating within the computational constraints required for clinical deployment. The specific objectives of this work are as follows: To design customized convolutional modules that capture the anisotropic anatomical features of spinal structures using directionally informed FE. To design customized convolutional modules that capture the anisotropic anatomical features of spinal structures using directionally informed FE. To implement AAM to enhance salient anatomical FS while attenuating background interference and imaging objects. To implement AAM to enhance salient anatomical FS while attenuating background interference and imaging objects. To formulate efficient Multi-Scale Feature Fusion Methods (MSFFM) that retain high-resolution boundary detail and broader anatomical context; To formulate efficient Multi-Scale Feature Fusion Methods (MSFFM) that retain high-resolution boundary detail and broader anatomical context; To achieve superior segmentation results relative to existing State-Of-The-Art (SOTA), as assessed by volumetric overlap, boundary precision, and computational performance metrics; and. To achieve superior segmentation results relative to existing State-Of-The-Art (SOTA), as assessed by volumetric overlap, boundary precision, and computational performance metrics; and. To rigorously evaluate the clinical utility of the proposed MRAS + FPN across heterogeneous patient cohorts and a range of pathological spinal conditions. To rigorously evaluate the clinical utility of the proposed MRAS + FPN across heterogeneous patient cohorts and a range of pathological spinal conditions. The main contributions of this work are fourfold: addressing critical gaps in current LS Segmentation. First, this study proposes a Multiscale Residual and Adaptive Spatial-Channel + Feature Pyramid Network (MRAS + FPN). This novel model integrates specialized Multiscale Residual Attention Blocks (MRAB) featuring Directionally Dilated Convolution Modules (DDCM) and dual spatial-channel Attention Mechanisms (AM), specifically designed to capture the anisotropic anatomical features of vertebral and disc structures. First, this study proposes a Multiscale Residual and Adaptive Spatial-Channel + Feature Pyramid Network (MRAS + FPN). This novel model integrates specialized Multiscale Residual Attention Blocks (MRAB) featuring Directionally Dilated Convolution Modules (DDCM) and dual spatial-channel Attention Mechanisms (AM), specifically designed to capture the anisotropic anatomical features of vertebral and disc structures. Second, this work introduces Adaptive Spatial-Channel Attention Blocks (ASCAB), which dynamically modulate Feature Selection (FS) across multiple pyramid levels, enabling enhanced boundary localisation and improved handling of anatomical variations via group normalisation and MSA. Second, this work introduces Adaptive Spatial-Channel Attention Blocks (ASCAB), which dynamically modulate Feature Selection (FS) across multiple pyramid levels, enabling enhanced boundary localisation and improved handling of anatomical variations via group normalisation and MSA. Third, this research develops a complete Cross-level Feature Gating (CFG) that optimises data flow between encoder-decoder pathways, ensuring the effective integration of MSA and spatial data while maintaining computational efficiency. Third, this research develops a complete Cross-level Feature Gating (CFG) that optimises data flow between encoder-decoder pathways, ensuring the effective integration of MSA and spatial data while maintaining computational efficiency. Fourth, this work provides extensive experimental validation indicating that MRAS-FPN achieves superior performance compared to recognized baselines, with an 87.8% Dice coefficient representing a 4.4% improvement over the best existing method (3D U-Net), while requiring 23% less training time and 15% less memory usage, effectively addressing the accuracy-efficiency trade-off that has limited clinical adoption of automated segmentation tools. These contributions collectively establish a new benchmark for LSS, bridging the gap between research successes and clinical requirements to provide a practical solution for automated spine analysis in clinical environments. Fourth, this work provides extensive experimental validation indicating that MRAS-FPN achieves superior performance compared to recognized baselines, with an 87.8% Dice coefficient representing a 4.4% improvement over the best existing method (3D U-Net), while requiring 23% less training time and 15% less memory usage, effectively addressing the accuracy-efficiency trade-off that has limited clinical adoption of automated segmentation tools. These contributions collectively establish a new benchmark for LSS, bridging the gap between research successes and clinical requirements to provide a practical solution for automated spine analysis in clinical environments. The rest of the article is organised as follows: Section “Related work” provides the related work; Section “Methodology” presents the methodology; Section “Experimental setup” describes the experimental setup; Section “Discussion” discusses the results; and Section “Conclusion” presents the conclusion. Related work Automated segmentation of LSS in MRI remains a challenging task due to high anatomical variability, the presence of imaging objects, and morphological changes associated with pathological environments. In response to these complexities, recent years have seen the emergence of increasingly sophisticated DL-based solutions. One notable improvement is ATM-Net29, which integrates anatomy-aware and text-guided MSA to align anatomical priors with visual features, enabling fine-grained segmentation of vertebrae and intervertebral discs. Similarly, the Attention LinkNet-152 model proposed by30 refines the traditional encoder-decoder model by incorporating deep residual encoding and attention-based skip connections, resulting in enhanced structural delineation across spinal regions31. adopted a diffusion-based semantic segmentation method, leveraging probabilistic refinement to achieve reliable differentiation of anatomical structures in patients experiencing lower back pain. In another hybrid model, SymTC32 combined parallel Transformer and CNN pathways to integrate global contextual cues with fine-grained local features, resulting in precise instance-level segmentation. Complementing these approaches33, introduced HCA-Net, a hierarchical context-aware attention network tailored for intervertebral disc labelling, while34 focused on automated quantification of the dural sac cross-sectional area using DL, presenting a clinically valuable tool for assessing spinal stenosis. In parallel with advances in task-specific segmentation, substantial research has focused on enhancing network design to improve FS and spatial decisions in Medical Image Segmentation (MIS). MSA has emerged as a key approach in this progression35. projected a Multi-Scale Feature Pyramid Fusion Network (MSFPFN) adapted explicitly for MIS, successfully integrating global anatomical coherence with fine-grained structural detail through the use of stacked pyramid features and AAM. Building on this foundation, MCBTNet36 introduced a hybrid model that combines multiple convolutional backbones with bi-level Transformer-based routing attention, significantly enhancing the model’s ability to capture long-range spatial dependencies. Another noteworthy contribution, R2U++37, extended the U-Net architecture into a multiscale recurrent residual network. By incorporating dense skip connections and deep supervision, R2U + + proved robust contextual encoding even under noisy or partially available data—a vital feature for clinical MRI scenarios. Residual Learning (RL) has significantly advanced deep segmentation by helping the retention and propagation of low-level spatial data across hierarchical layers, thereby improving model depth and trainability38. Invented a nested residual attention UNet + + architecture, which integrates channel-wise and spatial AM within residual pathways to enhance localization precision and mitigate boundary ambiguity. Extending this line of innovation39, introduced a dual-path synergy network that combines CNN with spatial self-attention Transformers, enabling synergistic interactions between local feature encoding and global contextual modeling. In a related effort40, incorporated visual Transformer modules into the U-Net, resulting in a hybrid segmentation model that improves generalization across multiple organ systems while maintaining computational efficiency. In addition to residual and MSA, AAM has emerged as a powerful tool for enhancing segmentation accuracy by dynamically modulating FS and mitigating background interference41. introduced the Swin Transformer into the UNETR, enabling hierarchical, window-based attention to capture global contextual dependencies in brain tumor segmentation from MRI scans. This method presented a favorable balance between contextual awareness and computational efficiency. In chest imaging42, employed a Convolutional Block Attention Module (CBAM) within a U-Net, achieving notable improvements in delineating lung boundaries. Building on these methods, DoubleU-NetPlus43 is an enhanced network that integrates dual encoder-decoder pathways with attention-based feature fusion and residual refinement, facilitating robust multi-scale learning across varied imaging modalities. Complementing this direction, Saeed et al. introduced 3D MFA, a multi-feature attention model for spine segmentation that augments MobileNetv3 with a Reverse Convolution Block Attention Module and Feature Pyramid Pooling. Three view-specific encoder networks trained independently on axial, coronal, and sagittal image planes generate concatenated 3D feature maps for final decoding, achieving Dice scores of 96.52% and 94.64% on the VerSe 2020 and VerSe 2019, at notably reduced computational cost via multi-stage network pruning. FPNs, initially developed for object detection, have increasingly been adapted and improved for MIS because they can aggregate hierarchical features across multiple spatial scales. Recent developments have produced several notable FPN variants tailored to specific medical domains. The SN-FPN44 employs nested feature hierarchies and self-attention mechanisms to support segmentation of complex tissue structures in digital pathology. In neuroimaging, enhanced FPN backbones have improved brain tumour segmentation by optimising the integration of top-down and bottom-up feature streams27. The FDUM-Net45 integrates FPN and U-Net to improve MIS precision in skin lesion analysis. Building on this direction, ISFPN46 combines FPN with Swin Transformers for ischemic stroke detection in brain CT images, achieving strong performance across lesion localisation and classification tasks. In the context of spine-specific imaging, Saeed et al. proposed an automated multi-scale feature fusion network for CT-based spine fracture segmentation, incorporating feature fusion modules, ASPP, and RCBAM for multi-scale spatial extraction and channel-wise refinement, establishing that hierarchical FPN-style feature aggregation yields reliable boundary localisation in the presence of anatomical variability. A parallel body of work has addressed CT-based and multimodal spine segmentation, offering complementary insights into feature extraction strategies relevant to spinal image analysis. Saeed et al. proposed MSFF, a multi-scale feature fusion model for MRI-based spine fracture segmentation combining ASPP, RCBAM, and a Residual Border Refinement Attention Block to generate discriminative multi-scale representations with sharpened boundary predictions. The same research group extended this framework to volumetric CT segmentation through 3D MRU-Net, which applies MobileNetv2 with residual blocks independently across axial, coronal, and sagittal views and concatenates the resulting feature maps for decoder-based segmentation, achieving competitive Dice scores on the VerSe datasets with reduced parameter overhead. For multi-task spinal analysis, Saeed et al. proposed a two-stage CT pipeline that combines CHASPPRAU-Net for spine segmentation with a 3D MRU-Net stage for vertebral labelling, validating the utility of cascaded architectures in integrated segmentation and recognition. Addressing multimodal clinical deployment, Saeed et al. presented a lumbar spine segmentation pipeline evaluated on both MRI and CT datasets, combining region detection, centroid annotation via SpatialConfigurationNet 3D, and HyperDenseNet-based segmentation further refined through gradient-sensitivity pruning, achieving Dice, IoU, and Hausdorff scores of 95.42%, 95.68%, and 4.51 mm. Collectively, these works highlight a persistent challenge across imaging modalities—capturing directional anatomical context and multi-scale semantic coherence without disproportionate computational cost—a gap the current study addresses through directionally dilated convolutions and cross-level adaptive gating within an MRI-specific FPN backbone. Taken together, these works highlight a typical trajectory in the field: the fusion of multiscale selections, attention-driven refinement, and reinforcement learning, frequently within a pyramid-structured backbone. However, while many existing models address specific segmentation challenges, few simultaneously optimise for boundary-level precision and cross-scale semantic coherence, particularly in the anatomically variable LS. The present study seeks to bridge this gap by proposing an MRAS-FPN, which integrates Directionally Dilated Convolutions, spatial-channel dual attention, and cross-level gating within an enhanced FPN backbone explicitly tailored for LS-MRI segmentation. Methodology Overview of the proposed model The proposed segmentation network, “MRAS-FPN”, is designed to address the dual challenge of achieving high anatomical precision and computational efficiency in lumbar spine MRI segmentation. The model builds on the initial FPN structure and augments it with two targeted attention-enhancement modules: the Multiscale Residual Attention Block (MRAB) and the Adaptive Spatial-Channel Attention Block (ASCAB). These components are purposefully integrated to enhance multiscale contextual feature learning, improve inter-level semantic consistency, and reduce data loss across the network hierarchy. The overall framework of MRAS-FPN follows an encoder-decoder method wherein the encoder comprises convolutional layers as to that FE increases semantic abstraction and decreases spatial resolution. These features are successively aggregated and upsampled using a top-down decoder path composed of feature maps to . The lateral connections between the encoder and decoder ensure the preservation of high-resolution localization data. At each stage of this pathway, the proposed MRAB modules are interposed to capture residual contextual data across multiple receptive fields and reinforce spatial and channel-wise attentional focus. The MRAB fuses low- and high-level features using a Directionally Dilated Convolution Module (DDCM), followed by a sequence of spatial and channel attention refinements and a Gated Fusion Method (GFM) to generate attention-enhanced feature maps. Following MRAB improvement, each decoded feature level is further modulated by ASCAB, which applies learnable attention across the spatial and channel axes in an adaptive, resource-efficient manner. This second stage of refinement ensures that only the most semantically relevant data is preserved for final segmentation prediction. The fused outputs are progressively upsampled and concatenated, culminating in a final prediction head that generates the binary segmentation map delineating vertebral structures and intervertebral discs. This model is precisely tailored to address the anatomical complexity of the lumbar spine, characterised by similar textures and blurred boundaries, by emphasizing cross-scale interactions and hierarchical attention control. The overall model is visualised in Fig. 1, which depicts the spatial hierarchy and module interactions within the FPN backbone. The proposed MRAS-FPN. Introduction to FPN The FPN serves as the network backbone of the proposed MRAS-FPN. The choice of FPN is motivated by its capacity to construct semantically rich feature hierarchies at multiple spatial resolutions, which is critical for segmenting anatomically complex structures such as the lumbar spine. The vertebrae and intervertebral discs vary significantly in scale, location, and shape across sagittal slices, displaying subtle boundaries that require a multiscale, high-resolution semantic representation. FPN achieves this goal by combining a bottom-up FEA pathway with a top-down feature fusion pathway, both of which are improved by lateral connections. The bottom-up pathway, typically instantiated using a residual convolutional encoder, processes the input image through multiple convolutional stages to generate a set of intermediate feature maps, Eq. (1). Where, is the feature map at stage , with spatial resolution progressively halved and channel depth ‘dl’ increasing use of the encoder depth. is the feature map at stage , with spatial resolution progressively halved and channel depth ‘dl’ increasing use of the encoder depth. These FS are increasingly abstract representations; however, lower-resolution layers suffer from a loss of fine spatial detail, which is crucial for precise segmentation. To recover high-resolution semantics, FPN introduces a top-down decoding mechanism that upsamples and merges features hierarchically. Each feature map ‘Pi’ in the pyramid, the value is computed by combining the upsampled higher-level feature with the corresponding lateral projection of , Eq. (2). where, is a convolution used to align channel dimensions. is a convolution used to align channel dimensions. is bilinear interpolation or transposed convolution that doubles the spatial resolution of . is bilinear interpolation or transposed convolution that doubles the spatial resolution of . The recursion is initialized by setting Eq. (3). After fusion, each ‘Pi’ is further refined with a convolution to disregard aliasing effects and enhance local discriminative capacity, Eq. (4). The resulting pyramid constitutes the core multiscale FS used in downstream segmentation layers. These outputs preserve localisation cues from shallow layers while leveraging the semantic depth of deeper layers, making FPN particularly suitable for structured anatomical segmentation. While FPNs have verified considerable success across several MIS tasks, the vanilla FPN exhibits notable problems when applied to anatomically complex scenarios, such as lumbar spine MRI. One primary limitation lies in the additive fusion used in the top-down pathway, which lacks the receptive field to capture the complex spatial dependencies and inter-channel relationships required to differentiate anatomically similar vertebral and disc structures. Also, the inherently static convolutional operations in the FPN backbone limit the network’s ability to adaptively emphasise diagnostically relevant anatomical features while suppressing background noise and irrelevant tissue regions. Also, the absence of residual connections in the fusion stages renders the architecture more vulnerable to gradient degradation and semantic loss, particularly in deeper layers of the network. To overcome these constraints, this study introduces two key improvements to the conventional FPN structure. The first is the Multiscale Residual Attention Block (MRAB), which is integrated at each pyramid stage to help context-sensitive feature learning. MRAB incorporates directionally dilated convolutions, residual summation paths, and a dual-AM that recalibrates spatial and channel-level FS. The second improvement is the ASCAB, applied after pyramid decoding to refine the fused features using group-wise normalization and MSA. Together, these modules transform the FPN backbone into a deeper, more anatomically aware, and dynamically adaptive model tailored to the unique challenges of LSS. Multiscale residual attention block (MRAB) Structure and integration in FPN The MRAB is introduced into the proposed MRAS-FPN to enhance the base FPN’s contextual sensitivity, MSA, and feature recalibration capabilities. Each MRAB module is embedded at an adjacent stage of the FPN’s top-down path, refining the decoded feature maps before they are propagated to the next pyramid level. This design ensures that each level-specific representation integrates spatially aligned high-resolution features with adaptively emphasized semantic cues. Structurally, the MRAB operates on two inputs: the high-level upsampled feature and the lateral encoder feature . The first step is to upsample to the resolution of , followed by an additive fusion to generate an intermediate FS, Eq. (5). where, Upsample is bilinear interpolation with a scale factor of 2. Upsample is bilinear interpolation with a scale factor of 2. This fused feature map is then processed using a Directionally Dilated Convolution Module (DDCM), which employs asymmetric dilated convolutions in orthogonal directions (i.e., horizontal and vertical) to extract orientation-aware contextual features. The outputs of this operation are concatenated and linearly predictable via a convolution, Eq. (6). where, ◊ Channel-wise concatenation. ◊ Channel-wise concatenation. ◊ Direction-aware spatial context that is particularly useful for detecting vertically and horizontally elongated structures such as vertebral borders and intervertebral clefts. ◊ Direction-aware spatial context that is particularly useful for detecting vertically and horizontally elongated structures such as vertebral borders and intervertebral clefts. To further refine the output, MRAB incorporates a dual AM composed of a Spatial Attention Block (SAB) and a Channel Attention Block (CAB), which are applied in sequence to . The SAB recalibrates feature intensities across the spatial domain by learning an attention mask , whereas the CAB generates a channel-wise importance vector . These refined activations are then fused using a CFG mechanism, which adaptively blends the attention-enhanced features with the original residual stream using a sigmoid-gated residual sum, Eq. (7). where, is element-wise multiplication. is element-wise multiplication. is the sigmoid activation function. is the sigmoid activation function. The term is the combined output of SAB and CAB applied to . This gating strategy ensures that only relevant attention-enhanced signals are retained while preserving the encoder’s critical low-level spatial fidelity. The final output is passed forward as the updated pyramid feature in the FPN’s top-down pathway, Eq. (8). The term is the combined output of SAB and CAB applied to . This gating strategy ensures that only relevant attention-enhanced signals are retained while preserving the encoder’s critical low-level spatial fidelity. The final output is passed forward as the updated pyramid feature in the FPN’s top-down pathway, Eq. (8). By inserting MRAB at each fusion stage in the pyramid, the MRAS-FPN can preserve multiscale contextual integrity, reinforce anatomical boundary cues, and prevent data loss that typically occurs in deep semantic hierarchies. The MRAB (Fig. 2) proves the flow of fused features through the DDCM, attention modules, and residual gating. MRAB. DDCM The DDCM serves as the introductory component (Fig. 3) of the MRAB, facilitating the extraction of spatial context from local and long-range anatomical orientations. The design rationale behind DDCM stems from the observation that vertebral bodies and intervertebral discs in lumbar spine MRI display prominent axial and sagittal elongation patterns. Traditional isotropic convolutional kernels lack sufficient directional sensitivity to capture such anatomical anisotropy, especially in the presence of degenerated or morphologically irregular structures. DDCM addresses this challenge by incorporating two orthogonal convolutional branches, each designed to encode spatial context with and without dilation along vertical and horizontal axes. The model of DDCM. Let as the intermediate fused feature map resulting from the top-down upsampled feature and lateral encoder feature at a given FPN level. This tensor is first passed through a conventional convolutional layer to unify the feature embedding, Eq. (9). ◊ Processed using two parallel branches: ◊ Processed using two parallel branches: Standard directional branch: This branch captures local spatial interactions using consecutive asymmetric convolutions, Eq. (10) Standard directional branch: This branch captures local spatial interactions using consecutive asymmetric convolutions, Eq. (10) where, The kernel emphasises vertical continuity, while the 1 × 3 kernel emphasises horizontal variation, forming a composite directional context within a compact receptive field. The kernel emphasises vertical continuity, while the 1 × 3 kernel emphasises horizontal variation, forming a composite directional context within a compact receptive field. Dilated directional branch: In parallel, a second branch applies dilated convolutions to extract long-range dependencies, Eq. (11). Dilated directional branch: In parallel, a second branch applies dilated convolutions to extract long-range dependencies, Eq. (11). where, DilatedConv(r) is a dilated convolution with dilation rate , typically set to 2 or 3. DilatedConv(r) is a dilated convolution with dilation rate , typically set to 2 or 3. This path enables extended context modelling without compromising spatial resolution. This path enables extended context modelling without compromising spatial resolution. The outputs of branches are summed to integrate local and non-local directionally encoded features, Eq. (12) Finally, a convolutional layer projects the fused representation into the desired embedding space, Eq. (13). The output tensor , it has hierarchical and orientation-aware features, enabling the network to detect subtle anatomical changes, such as disc-vertebral interfaces, posterior vertebral corners, and degenerative variations across sagittal slices. The DDCM module is lightweight, parameter-efficient, and free of pooling operations, ensuring minimal loss of spatial granularity. In contrast to traditional MSAs that use isotropic dilation or cascade networks, DDCM’s hybrid directional invention enables efficient yet expressive spatial encoding. Its integration into MRAB enhances the anatomical coherence and interpretability of feature maps across the FPN hierarchy, serving as a critical precursor to the spatial and channel AM discussed in the following sections. Spatial attention block (SAB) The SAB (Fig. 4) within the MRAB dynamically enhances the spatial saliency of semantic features by learning attention maps that highlight structurally relevant regions in the feature space. In lumbar spine MRI, where vertebrae and intervertebral discs frequently exhibit subtle boundaries and locally varying intensities, spatial attention disambiguates similar textures and reinforces position-sensitive activation patterns—especially around endplates, annular fissures, and the vertebral arch. SAB. SAB Takes Two Inputs: the contextually enriched feature derived from the DDCM and a modulation feature originating from lateral FPN streams or skip connections. SAB Takes Two Inputs: the contextually enriched feature derived from the DDCM and a modulation feature originating from lateral FPN streams or skip connections. From Fig. 4, both tensors are first channel-compressed via convolutions to reduce dimensional complexity, Eq. (14) Next, the spatial resolution of is aligned with that of using a learnable transposed convolution (deconvolution), Eq. (15) This upsampled tensor is then concatenated with , forming the composite attention seed, Eq. (16). To extract spatial saliency cues, this combined tensor is processed by two parallel convolutional branches: one estimating a spatial attention weight map , and the other generating an offset map which captures dynamic spatial displacement in both axes, Eq. (17). The final spatial attention operation applies a differentiable mapping function that repositions features in based on the learned offset map; it modulates the output using the attention weight (Eq. (18). This mapping operation is implemented using a soft attention formulation that interpolates features at spatially shifted coordinates and scales their contribution according to . It effectively allows the network to learn deformable spatial priors and adapt its focus to salient anatomical structures even under occlusion, distortion, or compression. Finally, the output is added residually to the input to preserve original spatial fidelity while embedding spatially calibrated emphasis, Eq. (19). This residual formulation ensures numerical stability during training and supports gradient flow during backpropagation. Significantly, the SAB operates without spatial pooling, thereby avoiding the resolution degradation commonly associated with global spatial AM. By selectively amplifying location-sensitive features and introducing deformable spatial adaptation, SAB plays a pivotal role in improving the anatomical alignment and localization precision of the MRAS-FPN, particularly in challenging cases with overlapping or morphologically degenerated spinal structures. CAB The CAB selectively amplifies or suppresses feature channels based on their global semantic relevance. In the context of LSS, different channels in the deep feature space frequently correspond to distinct anatomical or pathological cues—such as vertebral contours, disc spaces, or signal voids. However, not all channels contribute equally at every level of the pyramid. CAB addresses this by introducing a lightweight AM that models inter-channel dependencies through a global context-aware recalibration scheme. From Fig. 5, CAB operates on two inputs: the directionally enhanced feature map obtained from DDCM, and the modulation signal , which is typically derived from the lateral FPN stream at the corresponding level. First, a local encoding of the contextual feature is generated via a standard convolution applied to , Eq. (20). CAB. This descriptor is passed through a convolution to project the global context into a learnable channel attention vector , Eq. (21). where, is the sigmoid activation function that normalizes attention weights between 0 and 1. is the sigmoid activation function that normalizes attention weights between 0 and 1. The attention vector is then used to modulate the feature activations in using element-wise multiplication, Eq. (22) Simultaneously, the global semantic descriptor is derived from using Global Average Pooling (GAP), which compresses the spatial dimensions and generates a descriptor vector that reflects the average activation of each channel, Eq. (23) This operation allows the network to emphasize channels that encode highly relevant anatomical information while attenuating less discriminative filters. To maintain the integrity of the original spatial semantic distribution, CAB concludes with a residual connection, Eq. (24). This residual formulation ensures gradient stability during training and allows the attention operation to function as a semantic refinement rather than a complete overwrite. Moreover, since CAB relies solely on global pooling and pointwise convolution, it introduces negligible computational overhead and is well-suited for integration into memory-constrained architectures such as MRAS-FPN. By learning to prioritize feature channels in a data-driven, anatomically adaptive manner, the CAB significantly enhances the FS of the MRAB. When used in conjunction with the SAB, it enables the network to execute fine-grained, 2-D attention calibration—spatially and across channels—generating richer, more focused feature maps for downstream segmentation. Cross-level feature gating (CFG) The CFG mechanism serves as the final stage of the MRAB, enabling the model to selectively integrate context-aware semantic features from the attention-enhanced path with fine-grained spatial features from the backbone encoder. The principal motivation behind CFG is to dynamically regulate the contribution of deep semantic features vs. shallow spatial features at each pyramid level, ensuring that the final output preserves boundary localization while incorporating global anatomical context. From Fig. 6, the CFG module operates on two primary inputs: the output from the SpatialChannel Attention modules, denoted , and the earlier FS from the DDCM as . To enable element-wise interaction, the higher-level semantic map is first upsampled by a factor of 2, as shown in Eq. (25). The CFG. Simultaneously, is refined by a convolution to produce a lower-level spatial stream , Eq. (26). The two features are combined via an additive operation, as shown in Eq. (27). This fused representation is projected through a sigmoid-activated convolution to generate a gating , Eq. (28). The mask is then used to modulate the deep semantic representation , while simultaneously allowing residual information from to flow by Eq. (29) where, is element-wise multiplication. is element-wise multiplication. This gating operation ensures that highly confident semantic activations are preserved from the top-down decoder, while ambiguous or uncertain areas default to the high-resolution spatial signals. The resulting gated feature map undergoes a final transformation via a convolutional block, Eq. (30). Finally, to prepare the output for segmentation prediction, it is passed through a projection followed by upsampling to the original image resolution, Eq. (31) The output constitutes the refined segmentation map for the current level of the pyramid. When CFG modules are applied across all pyramid stages by the model achieves a balanced data flow between MSA and precise boundary cues. By introducing data-dependent, soft gating between hierarchical features, the CFG module addresses the long-standing challenge of reconciling detail preservation with semantic abstraction in deep MIS. It enables the MRAS-FPN to maintain high localisation accuracy without sacrificing global contextual understanding, which is crucial in clinically sensitive regions, such as the spinal canal and posterior vertebral corners. Adaptive spatial-channel attention block (ASCAB) The ASCAB (Fig. 7) is proposed as a terminal refinement mechanism within the MRAS-FPN, designed to modulate FS adaptively along spatial and channel dimensions. Unlike earlier attention operations within MRAB, which are embedded during MSA, ASCAB operates at the final stage of each decoder output To selectively enhance semantically significant regions while preserving feature diversity. This is particularly important in LSS, where subtle variations in texture and shape—such as disc compression, Schmorl’s nodes, or endplate irregularities —can easily be lost in deeper illustrations. ASCAB addresses this by applying a dual-branch attention mechanism that leverages global channel dependencies and local spatial context, with all operations conditioned on the feature map’s content. ASCAB. Let is the decoded feature map at a given pyramid level ‘i’. The ASCAB is composed of 3 structurally interlinked branches: (i) Channel attention, (ii) Spatial attention, and (iii) Identity-preserving residual path. Channel attention branch To determine the relative importance of each feature channel, ASCAB first constructs a global semantic descriptor using global average pooling, Eq. (32). This vector ‘Gc’, encoding the mean activation across spatial dimensions for each channel, is projected into a channel-wise modulation vector using a Fully Connected (FC) layer, followed by Group Normalization and sigmoid activation, Eq. (33) The resulting vector is a learned importance score for each channel, scaled to 0–1. It is then broadcast spatially and applied to the original feature map via element-wise multiplication, Eq. (34). This operation enables the suppression of noisy or low-utility channels and the amplification of semantically rich activations—crucial for emphasising key vertebral landmarks and discriminating adjacent anatomical units. Spatial attention branch While the channel attention pathway captures global semantics, the spatial attention branch identifies local, position-sensitive features. To this end, the input ‘M’ is passed through 3 independent convolutional layers to construct query , key , and value tensors, Eq. (35) where, , is a reduced embedding dimension. , is a reduced embedding dimension. An attention matrix is computed via matrix multiplication and SoftMax normalization, Eq. (36). The matrix ‘A’ captures inter-pixel similarity and is used to aggregate context via Eq. (37) Reshaping ‘R’ back to spatial form generates , which is normalized and regularized using Group Normalization and added back to the original input ‘M’, Eq. (38) This formulation enables ASCAB to adaptively reweight spatial locations based on intra-feature correlations and spatial relevance, which is critical for resolving closely packed vertebral units or disc-space ambiguity. Attention fusion and residual reintegration The outputs of the 2 branches are fused via additive aggregation (Eq. (39). To preserve original features and stabilize learning, ASCAB incorporates a residual identity connection, Eq. (40) This residual integration ensures that, while required activations are emphasised, baseline structural content is retained, preventing degradation of fine-edge details, particularly at transitional zones such as L5-S1 or T12-L1. In the MRAS-FPN, ASCAB is applied at each pyramid output level after MRAB refinement and upsampling; this placement allows it to serve as a final modulation layer before segmentation prediction, ensuring that semantic and structural representations are contextually filtered and anatomically focused. The dual-branch design proposes robustness to inter-class similarity and flexibility in inter-subject variability-a necessary feature when segmenting spines with pathological changes, vertebral fusion, or post-operative alterations. Loss function and optimization method The loss function in the proposed MRAS-FPN is designed to address the dual objective of volumetric segmentation accuracy and boundary localization precision, both of which are critical in LS image analysis. While accurate classification of vertebral and intervertebral disc regions is vital, precise delineation of limits is equally essential to support clinical applications such as surgical planning and pathology detection. To achieve this, a composite loss formulation is employed that integrates region-overlap-based metrics, pixel-level classification penalties, and edge-aware regularisation. This section presents the complete mathematical formulation of the loss function, the rationale for the selected weights, and the corresponding optimisation method. Composite loss function The total loss function is a weighted combination of the Dice loss , Binary Cross-Entropy loss , and an optional Boundary loss , Eq. (41) where, Scalar weights are assigned to each term based on its intended impact during training. Scalar weights are assigned to each term based on its intended impact during training. Dice loss: Dice loss is a region-based metric designed to maximize the overlap between the predicted segmentation map and the ground truth mask Y, which is particularly effective for dealing with class imbalance, Eq. (42)42 where, Dice loss: Dice loss is a region-based metric designed to maximize the overlap between the predicted segmentation map and the ground truth mask Y, which is particularly effective for dealing with class imbalance, Eq. (42)42 where, N is the sum of pixels. N is the sum of pixels. ε is a small constant added for numerical stability. ε is a small constant added for numerical stability. This term encourages global volumetric consistency and ensures high overlap with the target mask. (b)Binary cross-entropy loss (BCEL): To provide pixel-level supervision, the BCEL is used, which penalizes incorrect classification probabilities at each pixel location, Eq. (43)43 While Dice loss addresses overall shape alignment, BCEL reinforces local consistency and is particularly effective during the early stages of training. Binary cross-entropy loss (BCEL): To provide pixel-level supervision, the BCEL is used, which penalizes incorrect classification probabilities at each pixel location, Eq. (43)43 While Dice loss addresses overall shape alignment, BCEL reinforces local consistency and is particularly effective during the early stages of training. (c)Boundary-aware loss (BAL): To improve segmentation accuracy at region edges, an optional BAL term is introduced. This term compares the gradients of the predicted and true masks using an -norm formulation, Eq. (44)44 where, Boundary-aware loss (BAL): To improve segmentation accuracy at region edges, an optional BAL term is introduced. This term compares the gradients of the predicted and true masks using an -norm formulation, Eq. (44)44 where, The number of boundary pixels. The number of boundary pixels. is the spatial gradient operator used to extract edge maps. is the spatial gradient operator used to extract edge maps. This loss enhances contour fidelity and reduces edge-level misclassifications in structurally ambiguous regions. This loss enhances contour fidelity and reduces edge-level misclassifications in structurally ambiguous regions. (d)Weighting method: In the implementation of the composite loss function, the weighting parameters are selected based on empirical validation and network priorities. Equal weights are assigned to the Dice and BCE components, i.e., , reflecting the requirement to optimize for simultaneous overlap quality and pixel-level classification accuracy. The boundary-aware term is treated as a regularization component and is therefore scaled with a smaller weight of . This configuration ensures that boundary sharpness is enhanced without overpowering the primary learning signal. These coefficients are fixed across all experiments unless explicitly varied during ablation studies. Weighting method: In the implementation of the composite loss function, the weighting parameters are selected based on empirical validation and network priorities. Equal weights are assigned to the Dice and BCE components, i.e., , reflecting the requirement to optimize for simultaneous overlap quality and pixel-level classification accuracy. The boundary-aware term is treated as a regularization component and is therefore scaled with a smaller weight of . This configuration ensures that boundary sharpness is enhanced without overpowering the primary learning signal. These coefficients are fixed across all experiments unless explicitly varied during ablation studies. Optimization strategy The MRAS-FPN is trained end-to-end using the AdamW optimizer, which combines the adaptive learning rate behavior of Adam with decoupled weight decay. The training hyperparameters are set as follows: Initial learning rate: Initial learning rate: Weight decay: Weight decay: Scheduler: Cosine annealing with warm restarts every 20 epochs. Scheduler: Cosine annealing with warm restarts every 20 epochs. Batch size: 8. Batch size: 8. Epochs: 150. Epochs: 150. Gradient clipping: 5.0 to prevent exploding gradients in deep attention modules. Gradient clipping: 5.0 to prevent exploding gradients in deep attention modules. To improve model generalization and robustness, particularly given the limited size of annotated spine datasets, extensive data augmentation is applied during training. This includes horizontal and vertical flipping and small-angle rotations , gamma correction, histogram equalization (CLAHE), and random intensity jittering. Experimental setup Ethics statement This study was conducted in full compliance with institutional and national ethical standards governing research involving human subjects. This study utilized the publicly available MRSpineSeg Challenge dataset for computational analysis. All procedures performed in this study were tested in accordance with relevant research guidelines and regulations. Since the dataset consisted of anonymized retrospective medical images and did not involve direct human participation, patient recruitment, or clinical intervention, ethical approval and informed consent were not required. Datasets This study uses the MRSpineSeg dataset, which comprises T2-weighted volumetric lumbar spine MRI scans from 215 subjects. The data were collected from a local hospital and include healthy individuals and patients with a range of spinal pathologies, ensuring anatomical diversity and clinical significance. Each MRI volume is accompanied by expert-corrected manual segmentations of lumbar vertebrae and intervertebral discs, generated using ITK-SNAP. Up to 19 spinal structures are annotated per subject, with labels covering individual vertebrae, discs, and backgrounds across 20 classes. The in-plane pixel spacing ranges from 0.30 to 0.59 mm, and slice thicknesses range from 4.40 to 5.50 mm, reflecting the anisotropic resolution characteristic of clinical T2-weighted sagittal acquisitions. Image dimensions range from 512 × 512 to 1024 × 1024 pixels, with 12–18 slices per volume. MRAS-FPN processes each sagittal slice independently as a 2D input, preserving full in-plane spatial fidelity while avoiding the propagation of inter-slice partial volume artifacts across network layers. This per-slice design is appropriate for MRI sequences acquired with thick sagittal slices, where through-plane resolution is insufficient to support reliable volumetric feature aggregation. Interslice spatial context is not explicitly modeled, which represents a constraint of the current implementation relative to isotropic 3D acquisition protocols. To ensure robust evaluation, this study employs the dataset’s 5-fold cross-validation protocol, with each fold comprising 43 subjects. Data augmentation during training includes random rotations, elastic deformations, and contrast variations to improve model generalisation. For experimental purposes, the dataset was randomly divided into training (70%), validation (15%), and test (15%) splits, ensuring no subject overlap. All images were normalized using z-score normalization, and standard augmentations, including flipping, affine distortion, and intensity jitter, were applied during training to enhance model generalizability. Training configuration The proposed MRAS-FPN was implemented using PyTorch 2.0 and trained on a high-performance workstation equipped with an NVIDIA RTX 3090 GPU (24 GB VRAM). The training was conducted in mixed-precision mode using NVIDIA Apex AMP to optimize memory usage and computational speed. All experiments were performed under deterministic conditions, with fixed random seeds to ensure reproducibility. Each sagittal MRI slice was preprocessed (Fig. 8) using z-score normalisation to standardise intensity distributions across patients and imaging protocols. No spatial resizing was applied, and original voxel dimensions were preserved to maintain anatomical fidelity. To improve robustness to variability in spinal morphology and MRI acquisition conditions, a set of data augmentations was applied during training. These included random horizontal flipping, rotation within ± 15°, intensity jittering, elastic deformation, and CLAHE (Contrast-Limited Adaptive Histogram Equalization). Preprocessing steps. Optimization was performed using the AdamW optimizer, initialized with a learning rate of , a weight decay of , and a batch size of 8. A cosine annealing schedule with warm restarts was applied every 20 epochs to dynamically modulate the learning rate. The model was trained for 150 epochs, with early stopping employed if validation loss failed to improve for 15 consecutive epochs. A gradient clipping threshold of 5.0 was enforced to stabilize backpropagation using the deeper attention-enhanced modules. Model checkpoints were selected based on the best mean Dice coefficient on the validation set. All reported results correspond to the checkpoint with the highest validation performance across all anatomical labels. Table 1 provides the training configuration for the proposed model. Table 1Training configuration for MRAS-FPN.ParameterValue/descriptionModelPyTorch 2.0GPU hardwareNVIDIA RTX 3090 (24 GB)Precision modeMixed precision (NVIDIA Apex AMP)NormalizationZ-score (Zero Mean, Unit Variance)Data augmentationRandom Flip, Rotation (± 15°), Intensity Jitter, Elastic Deformation, CLAHEOptimizerAdamWInitial learning rate Weight decay Learning rate schedulerCosine annealing with warm restarts (Every 20 Epochs)Batch size8Epochs150Gradient clipping threshold5.0Early stopping patience15 epochs (Based on Validation Loss)Model checkpointing criterionThe best mean Dice score on the validation set Training configuration for MRAS-FPN. Evaluation system of measurement To commonly assess the segmentation performance of the proposed MRAS-FPN on LS-MRI, we employ a set of standard metrics that capture region-level accuracy and boundary precision. These metrics are computed per structure (vertebrae, intervertebral discs, and spinal canal), and the final scores are reported as mean values across all annotated classes, unless otherwise stated. Dice similarity coefficient (DSC) The DSC is used as the primary evaluation metric due to its sensitivity to region overlap and its widespread adoption in MIS tasks. It is defined as Eq. (45) where, is the ground truth binary mask. is the ground truth binary mask. is the predicted probability map (binarized using a threshold of 0.5). is the predicted probability map (binarized using a threshold of 0.5). N is the sum of voxels. DSC values range from 0 to 1, with higher values indicating better overlap between prediction and ground truth. N is the sum of voxels. DSC values range from 0 to 1, with higher values indicating better overlap between prediction and ground truth. HD95 To evaluate the spatial alignment between predicted and true boundaries, this study computes the HD95. This metric measures the worst-case surface difference while reducing sensitivity to outliers, Eq. (46) where, A, B are the sets of surface points from the predicted and ground truth masks. A, B are the sets of surface points from the predicted and ground truth masks. Lower HD95 values indicate better boundary agreement and are crucial for evaluating segmentation reliability in clinical contexts where anatomical precision is vital. Lower HD95 values indicate better boundary agreement and are crucial for evaluating segmentation reliability in clinical contexts where anatomical precision is vital. Average symmetric surface distance (ASSD) This study also reports the ASSD to capture the average deviation between predicted and actual structure boundaries, Eq. (47) ASSD is particularly informative for assessing fine-grained structural misalignments that may not be captured by volumetric overlap metrics alone. Precision and recall To measure classification performance at the pixel level, precision and recall are computed as: where, TP, FP, and FN are the number of True Positives, False Positives, and False Negatives. TP, FP, and FN are the number of True Positives, False Positives, and False Negatives. These metrics are crucial for measuring under- or over-segmentation tendencies in highly imbalanced structures, such as small discs or compressed vertebrae. Mean intersection over union (mIoU) To further evaluate the per-class segmentation quality across all anatomical labels, we compute the mIoU. This metric reflects the average spatial agreement between the predicted and ground truth regions across all structures, Eq. (49) where, C is the number of classes. C is the number of classes. are the predicted and true masks for class ‘c’. are the predicted and true masks for class ‘c’. mIoU provides a stricter assessment than DSC, especially under conditions of structural imbalance or partial occlusion. mIoU provides a stricter assessment than DSC, especially under conditions of structural imbalance or partial occlusion. It is reported as a percentage, with higher values indicating greater accuracy. Qualitative comparison of segmentation results Figure 9 presents the qualitative segmentation results on four representative cases from the SPIDER-MRI test set. Each row corresponds to a sagittal lumbar spine MRI slice, followed by the corresponding ground truth and predictions from 4 challenging segmentation models and the proposed model. The anatomical labels include vertebral bodies from T10 to L5, intervertebral discs, and sacral transition regions, with a class-specific color legend visible on the right. Visualisation results between different methods. The second column in each row provides the manually annotated ground truth, which serves as the reference standard. The third column displays the segmentation result generated by the proposed MRAS-FPN. Visually, MRAS-FPN proves the highest consistency with the ground truth in terms of anatomical structure, label continuity, and spatial precision. Vertebral boundaries are well-defined, intervertebral disc regions are cleanly segmented, and no class confusion is observed along the spinal canal or at transitional boundaries. In contrast, the output of DeepLab v3 + samples localized boundary errors and region fragmentation, particularly near the L4/L5 and L5/S1 intervertebral disc spaces. 3D U-Net achieves relatively better anatomical coherence but suffers from over-smoothing and blurred vertebra-disc transitions. 2D U-Net produces acceptable results in mid-lumbar segments but displays notable mislabelling and segmentation collapse at the lower spine, especially under anatomical variation or pathology-induced distortion. These visual differences highlight the effectiveness of the MRAS-FPN’s dual-attention design and multiscale residual refinement in preserving the global structure and achieving fine-grained boundary accuracy. The qualitative improvements align with the observed gains in Dice similarity, mIoU, and HD95 reported in the quantitative evaluation. Quantitative results Performance comparison of overall segmentation The comprehensive evaluation results (Table 2) validate the superior performance of the proposed MRAS-FPN across all standard segmentation metrics when compared to established baseline methods on the MRSpineSeg dataset. MRAS-FPN achieves a remarkable DSC of 87.8 ± 1.3%, representing significant improvements of 9.9% over 2D U-Net (77.9%), 4.4% over 3D U-Net (83.4%), and 6.3% over DeepLab v3+ (81.5%). The method also achieves an exceptional mIoU of 78.2 ± 2.1%, outperforming all baselines by substantial margins of 12.0%, 4.8%, and 10.1%. The relatively low standard deviations across all metrics indicate consistent performance across different patient cases and anatomical variations, a crucial factor for clinical reliability. Particularly impressive are the boundary-related improvements, with HD95 of 3.18 ± 0.8 mm showing reductions of 34.0%, 19.9%, and 23.4% compared to the baseline methods, while ASSD of 1.32 ± 0.5 mm validates improvements of 38.6%, 25.0%, and 30.2%. Table 2Overall segmentation performance comparison on MRSpineSeg dataset.MethodDSC (%)mIoU (%)HD95 (mm)ASSD (mm)Precision (%)Recall (%)2D U-Net77.9 ± 1.566.2 ± 3.44.82 ± 1.22.15 ± 0.879.3 ± 2.176.8 ± 2.33D U-Net83.4 ± 2.573.4 ± 2.73.97 ± 1.11.76 ± 0.684.9 ± 2.282.1 ± 2.4DeepLab v3+81.5 ± 1.268.1 ± 2.94.15 ± 0.91.89 ± 0.782.7 ± 1.880.4 ± 1.9MRAS-FPN (Ours)87.8 ± 1.3*78.2 ± 2.1*3.18 ± 0.8*1.32 ± 0.5*89.4 ± 1.6*86.3 ± 1.7* Overall segmentation performance comparison on MRSpineSeg dataset. The balanced precision (89.4 ± 1.6%) and recall (86.3 ± 1.7%) values validate the robustness of MRAS-FPN, signifying an excellent ability to minimise FP and FN predictions—critical requirements in MIS. All performance improvements are statistically significant (p < 0.05, paired t-test), confirming that the experimental gains are not due to random variation. From a clinical perspective, the 4.4% improvement in DSC over the best baseline (3D U-Net) translates into more accurate volume estimates for surgical planning, while the 19.9% reduction in HD95 ensures more precise boundary localisation for classifying pathological changes. The performance gap between MRAS-FPN and existing methods is most definite in boundary-related metrics, highlighting the use of the proposed MSA and RL components in capturing fine-grained structural details. Among the baseline methods, 3D U-Net performs best overall due to its ability to capture volumetric context. At the same time, 2D U-Net performs worst due to limited spatial information, and DeepLab v3 + performs in the middle, with its atrous convolutions providing some multi-scale capability but lacking the sophisticated AM that makes MRAS-FPN particularly effective for LSS tasks. Individual vertebra segmentation performance The individual vertebra segmentation results in Tables 3 and 4 reveal significant anatomical-region-specific performance variations and confirm MRAS-FPN’s superior capability across all vertebral levels. The most pronounced improvements are observed in the challenging upper thoracic regions (T9–T11), where baseline methods are consistently limited. For T9 segmentation, both 2D and 3D U-Nets result in complete segmentation failures (0.00% DSC), whereas DeepLab v3 + achieves only 26.83% DSC. MRAS-FPN reaches 38.74% DSC, a result that, while modest in absolute terms, represents a meaningful advance in this highly challenging anatomical region. Similarly, for T10, MRAS-FPN achieves 64.85% DSC compared to DeepLab v3+ (49.40%), 3D U-Net (32.72%), and 2D U-Net (28.45%), signifying the method’s greater capacity to capture subtle anatomical features with limited training data and high morphological variability. Table 3Individual vertebra segmentation performance (DSC %).MethodT9T10T11T12L1L2L3L4L5SMean2D U-Net0.00 ± 0.0028.45 ± 32.1868.92 ± 25.4174.85 ± 19.7276.23 ± 18.4477.18 ± 15.2977.84 ± 14.9679.67 ± 13.8282.15 ± 10.1883.47 ± 3.2164.88 ± 15.323D U-Net0.00 ± 0.0032.72 ± 35.8572.17 ± 27.8778.23 ± 21.8579.91 ± 20.2980.42 ± 16.9980.94 ± 16.6882.86 ± 15.3685.43 ± 11.3186.99 ± 3.5567.97 ± 17.58DeepLab v3+26.83 ± 32.3249.40 ± 38.0079.67 ± 22.2586.10 ± 12.6287.56 ± 9.6086.70 ± 11.8286.65 ± 12.5186.80 ± 10.8886.55 ± 8.6186.52 ± 3.0876.68 ± 16.17MRAS-FPN (Ours)38.74 ± 28.92*64.85 ± 34.21*85.42 ± 19.84*91.73 ± 11.24*92.89 ± 8.57*92.18 ± 10.56*92.67 ± 11.18*92.41 ± 9.73*91.84 ± 7.69*91.25 ± 2.74*84.40 ± 14.47* Individual vertebra segmentation performance (DSC %). Table 4Individual vertebra segmentation performance (mIoU %).MethodT9T10T11T12L1L2L3L4L5SMean2D U-Net0.00 ± 0.0019.84 ± 24.7357.28 ± 21.4563.89 ± 16.8465.47 ± 15.7266.84 ± 13.1867.53 ± 12.9469.72 ± 11.9573.24 ± 8.8475.18 ± 2.8955.90 ± 12.853D U-Net0.00 ± 0.0022.89 ± 28.3462.15 ± 23.4768.42 ± 18.8470.18 ± 17.4571.23 ± 14.6871.89 ± 14.4173.67 ± 13.2576.12 ± 9.8477.84 ± 3.2859.44 ± 14.36DeepLab v3+17.92 ± 26.1837.84 ± 30.8769.47 ± 19.2377.58 ± 10.9479.01 ± 8.3278.42 ± 10.2478.38 ± 10.8478.51 ± 9.4378.31 ± 7.4678.37 ± 2.6767.38 ± 13.62MRAS-FPN (Ours)26.73 ± 22.94*52.67 ± 27.89*77.24 ± 17.18*86.42 ± 9.73*88.19 ± 7.41*87.28 ± 9.15*87.89 ± 9.69*87.62 ± 8.44*86.94 ± 6.67*86.74 ± 2.38*76.77 ± 12.15* Individual vertebra segmentation performance (mIoU %). The limited segmentation performance at T9 and the upper thoracic boundary is attributable to three compounding factors. First, the MRSpineSeg dataset is acquired with a lumbar-focused field of view, and T9 lies at the uppermost extent of the sagittal acquisition window. As a result, T9 is only partially visible or absent in a subset of the 215 volumes, reducing the number of complete and annotated training examples available for this level. Second, the inherent morphological variability at the thoracolumbar junction is substantially greater than at mid-lumbar levels; vertebral bodies at T9–T11 articulate with the posterior rib cage, display transitional morphology between thoracic and lumbar forms, and exhibit greater inter-subject variability in size and orientation. Third, the intensity gradients between adjacent structures at this junction are subtler in T2-weighted sagittal acquisitions, making boundary discrimination more demanding for any 2D per-slice segmentation model. The complete failure of both U-Net variants at T9 confirms that this is a dataset-level constraint rather than a model-specific limitation. For clinical deployment at the thoracolumbar junction, dedicated acquisition protocols that ensure full T9 coverage and supplementary training data from this region would be necessary to achieve diagnostically reliable performance. The performance progressively improves toward the lumbar region, with MRAS-FPN achieving excellent DSC values above 90% for all lumbar vertebrae (L1–L5), representing improvements of 5–13% over the best baseline methods. The mIoU results further confirm these findings, with MRAS-FPN achieving the highest performance across all vertebral levels and producing particularly notable improvements in the lumbar spine region, where clinical accuracy is most critical. For the clinically meaningful L1–L5 vertebrae, MRAS-FPN achieves mIoU values ranging from 86.74% to 88.19%, representing 8–10% improvements over DeepLab v3 + and 15–18% improvements over the U-Net variants. The reduced standard deviations in MRAS-FPN results, particularly evident in L4 (9.73% vs. 15.36% for 3D U-Net), indicate more consistent performance across different patient anatomies and pathological conditions. The sacral region (S) also demonstrates robust performance at 91.25% DSC and 86.74% mIoU, confirming the method’s ability to handle the transition zone between the lumbar and sacral regions. Overall, MRAS-FPN achieves a mean vertebra segmentation performance of 84.40% DSC and 76.77% mIoU, representing improvements of 7.72% and 9.39% over the best baseline (DeepLab v3+), with all improvements being statistically significant and clinically meaningful for applications requiring precise vertebral boundary delineation. Individual intervertebral disc (IIVD) segmentation performance The IIVD segmentation results in Tables 5 and 6 validate MRAS-FPN’s exceptional performance in segmenting these anatomically complex structures, which present unique challenges due to their relatively thin profiles and variable signal characteristics on T2-weighted MRI. Similar to vertebra segmentation, the most dramatic improvements are experimental in the challenging upper thoracic regions, where MRAS-FPN achieves 48.26% DSC for T9/T10 compared to DeepLab v3+ (35.53%) and complete failure by both U-Net variants (0.00%). For T10/T11, MRAS-FPN reaches 76.18% DSC, representing a 10.79% improvement over DeepLab v3 + and substantial gains over 3D U-Net (42.60%) and 2D U-Net (38.72%). The performance consistently improves toward the lumbar region, with MRAS-FPN achieving outstanding DSC values exceeding 90% for all lumbar IVD (T12/L1 by L5/S), signifying improvements of 5–8% over DeepLab v3 + and 9–12% over the U-Net variants. The clinically critical L1/L2 to L4/L5 segments show particularly robust performance, with DSC values ranging from 91.47% to 93.56%, signifying highly accurate delineation of disc boundaries, which are essential for pathology testing and surgical planning. Table 5Individual IVD segmentation performance (DSC %).MethodT9/T10T10/T11T11/T12T12/L1L1/L2L2/L3L3/L4L4/L5L5/SMean2D U-Net0.00 ± 0.0038.72 ± 33.2472.84 ± 22.6776.89 ± 20.7377.58 ± 18.5977.12 ± 17.6479.84 ± 15.0880.17 ± 13.5281.34 ± 9.9564.94 ± 16.823D U-Net0.00 ± 0.0042.60 ± 36.8176.70 ± 25.2480.46 ± 23.0181.17 ± 20.6680.77 ± 19.6083.21 ± 16.7583.43 ± 15.0284.70 ± 11.0568.12 ± 18.68DeepLab v3+35.53 ± 39.8565.39 ± 34.7684.04 ± 17.5487.07 ± 13.8687.84 ± 12.8287.09 ± 14.1687.48 ± 13.3185.60 ± 12.3286.20 ± 7.3878.47 ± 17.33MRAS-FPN (Ours)48.26 ± 35.64*76.18 ± 31.24*89.85 ± 15.67*92.74 ± 12.39*93.56 ± 11.47*92.89 ± 12.67*93.21 ± 11.89*91.47 ± 11.04*92.13 ± 6.60*85.59 ± 15.40* Individual IVD segmentation performance (DSC %). Table 6Individual IVD segmentation performance (mIoU %).MethodT9/T10T10/T11T11/T12T12/L1L1/L2L2/L3L3/L4L4/L5L5/SMean2D U-Net0.00 ± 0.0026.74 ± 26.8962.84 ± 19.3767.23 ± 17.8468.14 ± 16.0267.58 ± 15.2170.38 ± 13.0470.76 ± 11.6972.47 ± 8.6256.24 ± 14.303D U-Net0.00 ± 0.0029.18 ± 30.4766.89 ± 21.7371.84 ± 19.8472.73 ± 17.8972.21 ± 16.9775.29 ± 14.5375.67 ± 13.0177.42 ± 9.5960.14 ± 16.00DeepLab v3+24.73 ± 34.2653.42 ± 29.6476.84 ± 15.1880.21 ± 12.0481.37 ± 11.1280.54 ± 12.2880.98 ± 11.5478.92 ± 10.6979.67 ± 6.4070.74 ± 14.79MRAS-FPN (Ours)33.47 ± 30.58*64.89 ± 26.74*83.67 ± 13.58*88.42 ± 10.76*89.73 ± 9.94*88.94 ± 10.99*89.34 ± 10.31*86.89 ± 9.59*87.84 ± 5.72*79.24 ± 13.13**Statistically significant improvement over all baselines (p < 0.05, paired t-test). Individual IVD segmentation performance (mIoU %). *Statistically significant improvement over all baselines (p < 0.05, paired t-test). The mIoU results for IVD segmentation further validate the superior performance of MRAS-FPN, achieving a mean of 79.24% compared to 70.74% for DeepLab v3+, 60.14% for 3D U-Net, and 56.24% for 2D U-Net, representing improvements of 8.50%, 19.10%, and 23.00%. The reduced Standard Deviations (SD) across all IVD levels, particularly evident in the lumbar segments (e.g., L1/L2: 11.47% vs. 12.82% for DeepLab v3+), specify more consistent and reliable segmentation performance across diverse patient populations and disc conditions. Notably, MRAS-FPN maintains excellent performance even in the challenging L5/S transitional zone (92.13% DSC, 87.84% mIoU), where anatomical variations and pathological changes are common. The overall mean IVD segmentation performance of 85.59% DSC represents a 7.12% improvement over the best baseline method, with all improvements being statistically significant and clinically meaningful for applications requiring precise disc boundary identification, such as disc herniation test, degenerative change quantification, and surgical intervention planning. Ablation studies To validate the contribution of each component in the proposed MRAS-FPN, we conducted complete ablation studies. These tests systematically evaluate the impact of individual modules and design selections on segmentation performance. Component-wise contribution analysis The comprehensive ablation study results in Table 7 provide critical insights into the individual contributions of each proposed component within the MRAS-FPN, signifying a systematic and additive improvement in performance with the addition of each module. The baseline FPN establishes a solid basis with 80.2% DSC and 69.4% mIoU. However, the addition of the MRAB provides the most substantial single improvement, boosting performance by 3.9% DSC and 4.4% mIoU to 84.1% and 73.8%, validating the effectiveness of directionally dilated convolutions and dual AM in capturing anatomical features across multiple scales. The ASCAB autonomously contributes a 2.1% DSC improvement (82.3%) when added to the baseline FPN, while the combination of MRAB and ASCAB without CFG achieves 85.9% DSC and 75.6% mIoU, representing a cumulative 5.7% DSC improvement over baseline. The inclusion of the complete MRAS-FPN + CFG achieves the highest performance of 87.8% DSC and 78.2% mIoU, representing an additional 1.9% improvement and signifying CFG’s key role in optimally integrating multi-scale features. The boundary-related metrics show consistent improvements across configurations, with the full MRAS-FPN achieving HD95 of 3.18 mm and ASSD of 1.32 mm—reductions of 24.5% and 33.0%, compared to baseline FPN. While computational overhead increases gradually (From 23.5 M to 33.8 M parameters and 45.2G to 58.9G FLOPs), the 44% parameter increase yields a substantial 7.6% DSC improvement, and the reduced standard deviations in the complete configuration (± 1.3% vs. ± 1.8% for baseline) indicate improved training stability and more consistent performance, validating the synergistic effect of all proposed components working together. Table 7Ablation study: component contribution.ConfigurationDSC (%)mIoU (%)HD95 (mm)ASSD (mm)Parameters (M)FLOPs (G)Baseline FPN80.2 ± 1.869.4 ± 2.94.21 ± 1.01.97 ± 0.823.545.2FPN + MRAB84.1 ± 1.573.8 ± 2.53.78 ± 0.91.68 ± 0.728.752.1FPN + ASCAB82.3 ± 1.671.2 ± 2.73.96 ± 0.91.84 ± 0.725.947.8FPN + MRAB + ASCAB (w/o CFG)85.9 ± 1.375.6 ± 2.33.52 ± 0.81.51 ± 0.631.255.6MRAS-FPN (Full)87.8 ± 1.378.2 ± 2.13.18 ± 0.81.32 ± 0.533.858.9 Ablation study: component contribution. Component analysis of MRAB The detailed ablation analysis of MRAB components in Table 8 reveals the individual contributions and synergistic effects of the DDCM, SAB, and CAB within the MSA. The removal of DDCM results in the most significant performance degradation, with DSC dropping to 82.7% and mIoU to 72.1%, representing losses of 1.4% and 1.7%, compared to the full MRAB, highlighting the critical role of directionally aware context extraction in capturing the anisotropic anatomical structures of the lumbar spine. Removing either the SAB or the CAB individually results in minor but notable performance decreases (83.1% and 83.4% DSC), with CAB removal having slightly less impact, signifying that spatial attention provides more critical anatomical localisation cues than channel attention for this task. The pairwise combinations validate the complementary nature of these components, with DDCM + CAB achieving slightly better performance (84.0% DSC, 73.7% mIoU) than DDCM + SAB (83.8% DSC, 73.5% mIoU), indicating that the combination of directional context extraction and channel-wise feature recalibration is particularly effective for spine segmentation. The inference time analysis shows a progressive increase from 168 ms without DDCM to 185 ms for the full MRAB, representing only a 10% computational overhead while achieving the optimal performance of 84.1% DSC and 73.8% mIoU, indicating that the complete MRAB configuration provides the best balance between segmentation accuracy and computational efficiency, with all components working synergistically to capture multi-scale anatomical features and attention-guided feature refinement necessary for precise LSS delineation. Table 8Ablation study: MRAB components.MRAB configurationDSC (%)mIoU (%)HD95 (mm)Inference time (ms)Without DDCM82.7 ± 1.472.1 ± 2.63.89 ± 0.9168 ± 9Without SAB83.1 ± 1.572.8 ± 2.53.82 ± 0.9172 ± 8Without CAB83.4 ± 1.473.1 ± 2.43.76 ± 0.8175 ± 9DDCM + SAB only83.8 ± 1.373.5 ± 2.33.71 ± 0.8182 ± 10DDCM + CAB only84.0 ± 1.473.7 ± 2.43.69 ± 0.8179 ± 9Full MRAB84.1 ± 1.573.8 ± 2.53.78 ± 0.9185 ± 10 Ablation study: MRAB components. AM variants The AM comparison in Table 9 proves the superior performance of the proposed ASCAB over conventional attention methods, supporting the design selections made for LSS. The baseline configuration without attention achieves 80.2% DSC and 69.4% mIoU with 23.5 M parameters, establishing the basis for comparison. Individual AMs show meaningful improvements, with spatial attention achieving 83.1% DSC (2.9% improvement) and channel attention achieving 82.6% DSC (2.4% improvement), indicating that spatial attention provides slightly greater benefit for anatomical structure localisation in spine imaging. The established AM, SE-Net and CBAM achieve competitive performance with DSCs of 83.8% and 84.4%, representing 3.6% and 4.2% improvements over the no-attention baseline, while requiring 28.2 and 29.1 million parameters, respectively. However, the proposed ASCAB achieves the highest performance, with 85.9% DSC and 75.6% mIoU, representing 5.7% and 1.5% improvements over the baseline and CBAM. While ASCAB requires the highest parameter count at 31.2 M (a 32.8% increase over the baseline), the performance gains validate this computational overhead, particularly considering the adaptive nature of the AM, which dynamically adjusts to anatomical variations. The superior performance of ASCAB validates the design philosophy of combining group normalisation with adaptive spatial-channel attention, explicitly tailored for MIS. Spatial localisation and feature-channel recalibration are crucial for accurately describing complex anatomical structures with subtle intensity variations and irregular boundaries in LS-MRI. Table 9Ablation study: AM.Attention configurationDSC (%)mIoU (%)Parameters (M)No attention80.2 ± 1.869.4 ± 2.923.5Spatial attention only83.1 ± 1.472.4 ± 2.627.8Channel attention only82.6 ± 1.571.9 ± 2.726.9SE-Net style attention83.8 ± 1.373.2 ± 2.428.2CBAM style attention84.4 ± 1.274.1 ± 2.329.1Adaptive spatial-channel (ASCAB)85.9 ± 1.375.6 ± 2.331.2 Ablation study: AM. Directionally dilated convolution analysis The DDCM variant analysis in Table 10 validates the proposed directional dilated convolution method over conventional convolutional approaches for capturing anatomical context in LSS. The standard 3 × 3 convolution baseline achieves 82.1% DSC and 71.2% mIoU, providing a foundation for comparing different convolutional designs. While dilated convolution with rate r = 2 shows modest improvement (82.9% DSC, + 0.8%) with 12% additional FLOPs, and separable convolution achieves slight gains (82.4% DSC, + 0.3%) with 8% computational savings, these methods fail to capture the directional anatomical FS of spinal structures adequately. Asymmetric convolution using sequential 3 × 1 and 1 × 3 kernels validates better performance (83.2% DSC, 72.3% mIoU) with only 5% computational overhead, highlighting the importance of directional FE for vertebral and disc boundary detection. The proposed Directional Dilated convolution achieves the highest performance of 84.1% DSC and 73.8% mIoU, representing improvements of 2.0% DSC and 2.6% mIoU over the standard convolution baseline and a 0.9% DSC improvement over asymmetric convolution, while requiring 15% additional FLOPs. This superior performance validates the design rationale of combining directional sensitivity with dilated receptive fields to capture local anatomical details and long-range spatial dependencies, which are essential for accurate spine segmentation. The 15% computational overhead is well compensated by substantial performance gains, particularly given that the directional expanded method effectively models the anisotropic nature of spinal anatomy, where vertebral bodies exhibit strong vertical orientation and intervertebral discs show horizontal elongation patterns, making this specialised convolutional design particularly suited for lumbar spine MRI analysis. Table 10Ablation study: DDCM variants.DDCM configurationDSC (%)mIoU (%)Computational CostStandard Conv 3 × 382.1 ± 1.671.2 ± 2.8BaselineDilated Conv (r = 2)82.9 ± 1.572.0 ± 2.6+ 12% FLOPsSeparable Conv82.4 ± 1.571.6 ± 2.7−8% FLOPsAsymmetric Conv (3 × 1, 1 × 3)83.2 ± 1.472.3 ± 2.5+ 5% FLOPsDirectional dilated (Ours)84.1 ± 1.573.8 ± 2.5+ 15% FLOPs Ablation study: DDCM variants. Loss function component analysis The loss function component analysis (Table 11) proves the critical importance of composite loss design for achieving optimal LSS performance, with each component contributing specific advantages to distinct segmentation quality features. Individual loss functions show distinct features, with Dice loss only achieving 85.2% DSC and substantial volumetric overlap but limited boundary precision (HD95: 3.67 mm, Boundary Quality: 6.8), while BCE loss only reaches 84.1% DSC with inferior boundary metrics (HD95: 3.89 mm, Boundary Quality: 6.2), indicating that neither method alone is sufficient for high-quality MIS. The combination of Dice and BCE losses with equal weighting (λ₁ = λ₂ = 1.0) significantly improves performance to 86.9% DSC and 77.1% mIoU while enhancing boundary quality to 7.4, signifying the complementary nature of region-based and pixel-wise supervision. The progressive addition of boundary-aware loss shows systematic improvements, with λ₃=0.1 achieving 87.3% DSC and notably improved boundary metrics (HD95: 3.32 mm, Boundary Quality: 8.1), while the optimal configuration with λ₃ = 0.2 reaches the highest performance of 87.8% DSC, 78.2% mIoU, and superior boundary precision (HD95: 3.18 mm, Boundary Quality: 8.6). However, increasing the boundary loss weight to λ₃ = 0.5 results in slight performance degradation (87.1% DSC, HD95: 3.28 mm), signifying that excessive emphasis on boundary preservation can compromise overall segmentation quality. The optimal weighting method, with λ₁ = λ₂ = 1.0 and λ₃ = 0.2, effectively balances volumetric accuracy, pixel-level classification, and boundary sharpness, which is particularly vital for LSS, where precise delineation of vertebral endplates and disc boundaries is vital for clinical diagnosis and treatment planning. This validates the composite loss design as a key contributor to MRAS-FPN’s superior performance. Table 11Ablation study: loss function components.Loss configurationDSC (%)mIoU (%)HD95 (mm)Boundary Quality*Dice loss only85.2 ± 1.475.8 ± 2.43.67 ± 0.96.8 ± 1.2BCE loss only84.1 ± 1.674.3 ± 2.63.89 ± 1.06.2 ± 1.4Dice + BCE (λ₁ = λ₂ = 1.0)86.9 ± 1.277.1 ± 2.23.45 ± 0.87.4 ± 1.1Dice + BCE + Boundary (λ₃ = 0.1)87.3 ± 1.377.6 ± 2.13.32 ± 0.78.1 ± 1.0Dice + BCE + Boundary (λ₃ = 0.2)87.8 ± 1.378.2 ± 2.13.18 ± 0.88.6 ± 0.9Dice + BCE + Boundary (λ₃ = 0.5)87.1 ± 1.477.4 ± 2.33.28 ± 0.88.3 ± 1.1*Boundary Quality measured as average edge F1-score across all structures. Ablation study: loss function components. *Boundary Quality measured as average edge F1-score across all structures. Training metrics The training convergence analysis (Table 12; Fig. 10) reveals distinct learning features across methods, with MRAS-FPN signifying superior training stability and sustained performance improvement throughout the extended training period. While baseline methods achieve early convergence—2D U-Net at 95 ± 8 epochs, DeepLab v3 + at 102 ± 9 epochs, and 3D U-Net at 110 ± 12 epochs—MRAS-FPN requires longer training (118 ± 7 epochs) but achieves substantially higher final performance with exceptional training stability (0.89) compared to the baselines (0.68–0.75). Table 12Training convergence analysis.MethodConvergence epochFinal accuracy (%)Peak dice (%)Peak mIoU (%)Training stability*2D U-Net95 ± 889.2 ± 1.877.9 ± 1.566.2 ± 3.40.723D U-Net110 ± 1292.8 ± 2.183.4 ± 2.573.4 ± 2.70.68DeepLab v3+102 ± 991.4 ± 1.681.5 ± 1.268.1 ± 2.90.75MRAS-FPN (Ours)118 ± 794.7 ± 1.487.8 ± 1.378.2 ± 2.10.89 Training convergence analysis. Training convergence analysis. The epoch-wise performance data (Table 13; Fig. 11) validate that all methods show similar initial learning rates through epoch 60. However, MRAS-FPN maintains consistent improvement beyond epoch 80, where baseline methods plateau, ultimately achieving 87.8% DSC and 78.2% mIoU compared to the best baseline (3D U-Net) at 83.4% DSC and 73.4% mIoU. Notably, MRAS-FPN exhibits the most consistent learning trajectory with minimal performance oscillations, as evidenced by the lowest convergence epoch variance (± 7) and the highest training stability score (0.89), indicating that the MSA and RL components facilitate stable gradient flow and effective feature learning throughout the network’s depth. The sustained performance improvement from epoch 100 to 150 (84.6% to 87.8% DSC) showcases the benefit of the cosine annealing schedule with warm restarts, allowing MRAS-FPN to continue refining attention weights and FS long after baseline methods have saturated, ultimately mitigating the longer training time with superior final performance and indicating the architecture’s capacity for continued learning and adaptation to the complex anatomical variations present in LS-MRI data. Table 13Epoch-wise performance data.Epoch2D U-Net3D U-NetDeepLab v3+MRAS-FPN (Ours)DicemIoUDicemIoUDicemIoUDicemIoU1052.338.748.935.251.137.849.636.42065.852.462.149.864.251.661.948.94072.158.969.756.271.358.170.857.36075.963.274.862.176.262.876.463.78077.265.179.166.878.865.481.269.110077.765.881.970.280.767.284.673.812077.966.183.172.881.367.986.976.414077.966.283.473.281.568.087.677.815077.966.283.473.481.568.187.878.2 Epoch-wise performance data. Epoch-wise analysis: (a) Dice score, (b) mIoU. The training efficiency analysis (Table 14; Fig. 12) confirms that MRAS-FPN achieves a well-calibrated balance between computational requirements and segmentation performance. With a training time of 22.1 ± 1.6 h, MRAS-FPN requires 17% more time than DeepLab v3+ (18.9 h) but 23% less than 3D U-Net (28.7 h), while memory usage of 15.8 ± 1.0 GB falls between DeepLab v3+ (12.4 GB) and 3D U-Net (18.6 GB), reflecting efficient memory management despite the added attention modules. GPU utilization of 89 ± 4% exceeds that of DeepLab v3+ (85 ± 4%) and 2D U-Net (78 ± 5%), approaching but not surpassing 3D U-Net (92 ± 3%), confirming effective hardware usage throughout training. Total FLOPs of 5.12T represent a 22% increase over DeepLab v3 + and a 24% reduction compared to 3D U-Net, signifying that MRAS-FPN delivers substantially higher segmentation accuracy at a computational cost lower than that of the volumetric baseline. Table 14Training efficiency analysis.MethodTraining time (Hours)Memory usage (GB)GPU utilization (%)Total FLOPs (T)Inference time (ms/slice)2D U-Net12.3 ± 0.88.2 ± 0.578 ± 52.848.6 ± 0.73D U-Net28.7 ± 2.118.6 ± 1.292 ± 36.7331.8 ± 2.4DeepLab v3+18.9 ± 1.412.4 ± 0.885 ± 44.2114.7 ± 1.1MRAS-FPN (Ours)22.1 ± 1.615.8 ± 1.089 ± 45.1218.3 ± 1.5Significant values are in bold. Training efficiency analysis. Significant values are in bold. Efficiency analysis. The per-slice inference time for the complete MRAS-FPN pipeline, measured under deterministic single-sample evaluation, is 18.3 ± 1.5 ms — 42.5% faster than 3D U-Net (31.8 ± 2.4 ms/slice) and only 3.6 ms slower than DeepLab v3+ (14.7 ± 1.1 ms/slice), a marginal overhead fully offset by the 6.3% DSC advantage. For a complete lumbar MRI volume comprising 12 to 18 sagittal slices, total inference time ranges from approximately 220 ms to 330 ms per volume. This throughput is well within the operational requirements of clinical decision support workflows, where per-case processing latency of several seconds is standard. Inference-time GPU memory consumption is substantially lower than the 15.8 GB training, as gradient computation and optimizer states are absent during inference, placing MRAS-FPN within reach of clinical workstations equipped with GPUs with 12 GB or more of VRAM. The validation loss analysis confirms MRAS-FPN’s superior learning dynamics, achieving the lowest validation loss of 0.162 at epoch 115 and maintaining stable convergence without early stopping, in contrast to baseline methods that reach their best validation performance earlier (2D U-Net: 0.259 at epoch 92; 3D U-Net: 0.206 at epoch 108; DeepLab v3+: 0.221 at epoch 99) but at higher final loss values. The epoch-wise validation loss trends (Table 15; Fig. 13) show MRAS-FPN consistently improving throughout training, with steady loss reduction from 0.786 at epoch 10 to 0.162 at epoch 115, followed by minimal fluctuation (0.162–0.167) in the final epochs, indicating robust convergence and excellent generalisation. In contrast, baseline methods exhibit earlier plateauing and slight degradation in the final epochs, whereas 2D U-Net increases from its best loss of 0.259 to 0.282 by epoch 150, signifying overfitting. The computational efficiency gains are particularly significant when considering the performance improvements: MRAS-FPN achieves 4.4% higher DSC than the best baseline (3D U-Net) while requiring 23% less training time and 15% less memory, resulting in a superior accuracy-to-computational-cost ratio that makes it highly suitable for practical implementation in clinical environments where performance and resource constraints are critical considerations. Table 15Validation loss curves and early stopping analysis.Epoch2D U-Net3D U-NetDeepLab v3+MRAS-FPN (proposed)100.7420.7980.7610.786200.5210.5780.5420.563300.3980.4450.4120.428400.3340.3670.3410.351500.2980.3190.2940.289600.2780.2840.2670.251700.2650.2510.2450.218800.2610.2270.2340.193920.2590.2120.2240.171990.2620.2080.2210.1651000.2640.2070.2220.1641080.2670.2060.2230.1631100.2690.2070.2240.1631150.2710.2090.2250.1621200.2730.2110.2260.1631300.2760.2140.2280.1641400.2790.2170.2300.1651500.2820.2200.2320.167Significant values are in bold. Validation loss curves and early stopping analysis. Significant values are in bold. Epoch vs. loss. Comparison with recent state-of-the-art methods To complement the baseline evaluation in Section “Quantitative results”, this subsection presents a structured comparison between MRAS-FPN and five task-specific lumbar spine segmentation methods published from 2024 onwards. Methods evaluated on MRSpineSeg are treated as direct comparisons; those evaluated on different datasets are included as indicative references and marked accordingly. All results are reproduced from the respective publications without modification. Among methods evaluated on the same MRSpineSeg dataset, MRAS-FPN achieves the highest DSC of 87.8%, surpassing ATM-Net29 by 6.08% points, nnU-NetV248 by 11.37% points, and U-BiSeNet47 by 21.77% points. The mIoU of 78.2% similarly exceeds ATM-Net (72.25%), nnU-NetV2 (67.50%), and U-BiSeNet (57.67%) by 5.95, 10.70, and 20.53% points. Boundary metrics for MRSpineSeg-evaluated literature methods are reported in pixel units; given variable in-plane pixel spacing across subjects, direct statistical conversion to millimeters is unreliable, and ordinal comparisons alone are appropriate within this group. MRAS-FPN’s HD95 of 3.18 mm and ASSD of 1.32 mm validate strong boundary precision, particularly relevant given that ATM-Net’s HD95 of 9.60 pixels would correspond to 2.88–5.66 mm across the dataset’s pixel spacing range. SymTC and SpineSegDiff report higher DSC values of 94.90% and 91.3%; these results stem from datasets with different subject counts, scanner protocols, and class definitions, and are not statistically comparable to MRSpineSeg performance. The consistent superiority of MRAS-FPN across all methods evaluated on the shared MRSpineSeg benchmark confirms the effectiveness of its directionally dilated convolutions and adaptive cross-level gating for fine-grained lumbar spine segmentation. Discussion Technical contributions and positioning within current literature The proposed MRAS-FPN marks a substantive development in automated LSS, presents targeted solutions to limitations identified in the existing literature, and introduces architectural innovations that jointly improve segmentation accuracy and computational efficiency. Rooted in the principles of FPN, MRAS-FPN extends this foundation by incorporating specialised components tailored to the anatomical and imaging complexities of the LS. In contrast to generalized FPN variants such as SN-FPN44 and ISFPN46, which are optimized for broader MIS, MRAS-FPN integrates directionally dilated convolutions that explicitly account for the anisotropic morphology of spinal structures. This targeted design addresses a critical limitation in prior MSAs, which predominantly relied on isotropic convolutional kernels and thus failed to differentiate between the vertical alignment of vertebral bodies and the horizontal elongation of intervertebral discs. Another distinguishing feature of the proposed method is the integration of dual spatial-channel AM into the MRAB and ASCAB. These components represent a refined evolution beyond standard AM, incorporating adaptive MSA that leverages group-wise normalization and residual learning to capture anatomical variability. While architectures such as Attention LinkNet-15230 and HCA-Net33 use attention layers, they lack the dynamic, structurally aware recalibration achieved by MRAS-FPN. Compared to CBAM-based developments within U-Net42, our method achieves superior quantitative performance, attaining a DSC of 87.8%, which falls within the typically reported range of 84–85%. Moreover, MRAS-FPN validates improved boundary delineation, achieving an HD95 of 3.18 mm, compared to over 4 mm reported in prior AM. Comparative analysis with state-of-the-art (SOTA) methods The experimental results presented in Section “Comparison with recent state-of-the-art methods” (Table 16) position MRAS-FPN as the strongest-performing method among all models evaluated on the MRSpineSeg dataset. Against the three methods sharing the same dataset and evaluation protocol, MRAS-FPN achieves a DSC of 87.8%, surpassing ATM-Net29 by 6.08% points, nnU-NetV248 by 11.37% points, and U-BiSeNet47 by 21.77% points. The corresponding mIoU improvements of 5.95%, 10.70%, and 20.53% over these same methods confirm that the performance advantage is consistent across both overlap-based metrics. These gains are particularly notable given that recent spine-specific models such as U-BiSeNet47 and ATM-Net29 already incorporate task-tailored architectural designs — multi-scale attention with BiSeNet fusion in the former and anatomy-aware text-guided multi-modal fusion in the latter. The fact that MRAS-FPN outperforms both with a purely vision-based convolutional attention architecture demonstrates that directionally dilated feature extraction and cross-level gating are highly effective design choices for this segmentation task without requiring the additional complexity of language alignment modules. Table 16Comparison with recent state-of-the-art methods.MethodYearDatasetDSC (%)mIoU (%)HD95ASDU-BiSeNet472024MRSpineSeg66.0357.6720.20 px†9.45 px†nnU-NetV2482024MRSpineSeg76.4367.5016.70 px†3.91 px†SymTC322024Private‡94.90 ± 2.10NR2.21 ± 1.19 mmNRSpineSegDiff312024SPIDER‡91.3NRNRNRATM-Net292025MRSpineSeg81.7272.259.60 px†2.15 px†MRAS-FPN (Proposed)2025MRSpineSeg87.8 ± 1.3*78.2 ± 2.1*3.18 ± 0.8 mm*1.32 ± 0.5 mm**Statistically significant improvement over all MRSpineSeg baselines (p < 0.05, paired t-test).† Boundary metrics reported in pixel units as published in the respective papers. In-plane pixel spacing in MRSpineSeg ranges from 0.30 to 0.59 mm across subjects, precluding reliable conversion to millimeters for cross-method comparison.‡ Evaluated on a different dataset; results are indicative only and not directly comparable to MRSpineSeg performance. SymTC uses a private 20-subject lumbar MRI dataset from the University of Miami. SpineSegDiff uses the SPIDER dataset comprising 218 subjects across four hospitals, with combined T1-weighted and T2-weighted MRI evaluation.Significant values are in bold. Comparison with recent state-of-the-art methods. *Statistically significant improvement over all MRSpineSeg baselines (p < 0.05, paired t-test). † Boundary metrics reported in pixel units as published in the respective papers. In-plane pixel spacing in MRSpineSeg ranges from 0.30 to 0.59 mm across subjects, precluding reliable conversion to millimeters for cross-method comparison. ‡ Evaluated on a different dataset; results are indicative only and not directly comparable to MRSpineSeg performance. SymTC uses a private 20-subject lumbar MRI dataset from the University of Miami. SpineSegDiff uses the SPIDER dataset comprising 218 subjects across four hospitals, with combined T1-weighted and T2-weighted MRI evaluation. Significant values are in bold. Compared with hybrid Transformer-CNN architectures such as SymTC32 and MCBTNet36, which integrate global context modelling with local convolutional encoding, MRAS-FPN achieves superior performance on the shared benchmark while maintaining substantially lower computational overhead. SymTC reports a DSC of 94.90% on a private 20-subject dataset, a figure that reflects favorable evaluation conditions rather than a directly comparable benchmark result; on the MRSpineSeg dataset, ATM-Net — SymTC’s closest architectural peer — achieves only 81.72% DSC, 6.08% points below MRAS-FPN. In terms of boundary precision, MRAS-FPN achieves a 19.9% reduction in HD95 relative to 3D U-Net and records an ASSD of 1.32 mm, improvements that substantially exceed the 5–10% HD95 reductions typically reported by attention-enhanced models such as R2U++37 and DoubleU-NetPlus43. The efficiency profile of MRAS-FPN further strengthens its clinical case. Training requires 22.1 h, representing a 23% reduction relative to 3D U-Net (28.7 h), while per-slice inference time of 18.3 ± 1.5 ms is 42.5% faster than 3D U-Net (31.8 ± 2.4 ms/slice) and only marginally slower than DeepLab v3+ (14.7 ± 1.1 ms/slice) — a difference of 3.6 ms per slice that is inconsequential in a clinical workflow context. Full-volume inference for 12 to 18 slices completes in approximately 220 to 330 ms, well within acceptable latency for offline diagnostic support. In contrast, Transformer-based models such as Swin UNETR41 and the hybrid visual Transformer U-Net40 require substantially greater memory and extended training durations, limiting their practical deployment in resource-constrained radiology environments. ATM-Net29, while achieving 81.72% DSC on MRSpineSeg, introduces the overhead of a natural language alignment pipeline; MRAS-FPN surpasses it by 6.08% DSC while operating entirely within the vision domain, offering a more computationally transparent and deployment-ready architecture. Clinical implications and practical impact The superior performance of MRAS-FPN translates into tangible clinical benefits, addressing key diagnostic and operational challenges in LSS. Notably, the enhanced boundary accuracy—reflected in an HD95 of 3.18 mm, compared to over 4 mm reported in existing methods—is particularly valuable in surgical planning contexts, where accurate delineation of vertebral endplates and intervertebral disc boundaries can directly affect the practicality and results of the method. This degree of precision closely aligns with the inter-observer variability observed in expert manual segmentations, highlighting the potential of MRAS-FPN as a dependable clinical decision-support tool. Moreover, the model validates consistent segmentation performance across the full range of vertebral levels (T9–S1), with DSC exceeding 90% in lumbar regions. This consistency addresses a common limitation in current methods, which frequently exhibit reduced accuracy in anatomically complex regions, such as the thoracolumbar junction. Equally significant is the method’s robustness in the presence of pathological variability. Unlike many existing segmentation networks, which are primarily evaluated on relatively normal anatomy, MRAS-FPN has been comprehensively validated on the MRSpineSeg dataset, which encompasses a broad spectrum of degenerative spine conditions. The model’s ability to maintain performance under such variability proves its practical relevance for deployment in real-world clinical settings, where anatomical deviations are standard. Notably, the reduced Standard Deviation (SD) across all key metrics (± 1.3% DSC compared to ± 2.5% in baseline methods) reflects greater reliability and predictability—critical traits for fostering clinical trust. This performance stability contrasts with probabilistic methods, such as diffusion-based segmentation (Monzon et al., 2024), which frequently exhibit greater variability due to their inherent stochastic refinement. Methodological innovations and generalizability The architectural innovations embedded within MRAS-FPN contribute several broadly applicable design principles that extend beyond the specific domain of LSS. The DDCM introduces a structured method for modeling anisotropic anatomical structures and validates potential utility for segmenting other elongated or directionally biased anatomical entities, such as vascular pathways, neural tracts, or cardiac chambers. Unlike standard dilated convolution methods—such as those used in nested residual attention UNet++ (Li et al., 2022)—DDCM incorporates directional sensitivity tailored to the morphological features of medical anatomy, representing a meaningful advance in the spatial modelling capabilities of convolutional layers. Also, the CFG mechanism recommends a refined alternative to the additive or concatenative fusion methods commonly employed in conventional FPN. By enabling learned, data-adaptive integration of multi-scale features, CFG enhances representational richness and contextual coherence, with probable applicability to a range of MIS tasks involving complex anatomical hierarchies. The ASCAB further addresses critical limitations of earlier attention modules by incorporating group normalization and residual summation, thereby improving training stability and cross-sample generalization. This design surpasses commonly used attention strategies, such as SE-Net and CBAM, by introducing dynamic, content-aware modulation that flexibly adapts to anatomical variability. The use of these components is verified by complete ablation studies that not only validate their contributions within the MRAS-FPN but also provide generalizable insights into the design of robust AM for MIS. These findings contribute to a broader understanding of network-level innovations in DL for healthcare applications. Limitations and future research directions Despite the significant improvements achieved by MRAS-FPN, several limitations remain that should be addressed in future work. The current evaluation is restricted to T2-weighted sagittal MRI sequences, whereas clinical protocols frequently include multiple imaging modalities such as T1-weighted and STIR sequences, which provide complementary anatomical information. Future research should consider extending MRAS-FPN to support multimodal MRI fusion and evaluate architectures such as ATM-Net29 that integrate multiple MRI sequences via cross-modal alignment. The model processes each sagittal slice independently as a 2D input, preserving in-plane spatial fidelity but not modeling inter-slice spatial context. Extension to isotropic high-resolution 3D acquisition protocols, where through-plane resolution is comparable to in-plane resolution, represents a natural next step that would enable full volumetric feature aggregation across the spinal axis. The model’s performance at the thoracolumbar junction, particularly at T9, reflects a dataset-level constraint arising from partial field-of-view coverage and the limited number of complete training examples at this level. Achieving diagnostically reliable segmentation at T9–T11 in clinical deployment would require dedicated acquisition protocols ensuring full thoracolumbar coverage and supplementary annotated data from this anatomical region. The composite loss function’s boundary weight λ₃ = 0.2 was optimized through systematic ablation within the T9–S1 lumbar range covered by MRSpineSeg; its sensitivity across other spinal regions, such as the cervical or thoracic spine, has not been evaluated and represents a parameter that would require reassessment when extending the model to broader anatomical coverage. The model’s robustness under complex pathological conditions, including severe spondylolisthesis, post-operative anatomy, and advanced spinal stenosis, also requires validation using dedicated datasets that capture this clinical variability, as MRSpineSeg predominantly reflects a degenerative disease spectrum and excludes extreme post-surgical cases. Further efficiency improvements are achievable through knowledge distillation or structured pruning, enabling faster inference on more constrained hardware without significant performance degradation. Incorporating federated learning could support collaborative model development across institutions without compromising patient data privacy, addressing the scarcity of large annotated spine MRI datasets. Adapting MRAS-FPN to other spinal regions or related tasks, such as spinal cord segmentation, nerve root delineation, or cervical spine analysis, would further establish the generalizability of the proposed directional convolution and adaptive attention design principles. Conclusion This study introduces MRAS-FPN, a DL designed to address the core challenge of achieving high segmentation accuracy while minimising computational cost in automated LS-MRI analysis. Evaluated on the MRSpineSeg dataset comprising 215 T2-weighted lumbar spine MRI volumes, MRAS-FPN achieves strong results, recording a DSC of 87.8 ± 1.3% and a mIoU of 78.2 ± 2.1%. These results echo the 4.4% and 4.8% improvements over the best-performing baseline (3D U-Net). Notably, boundary accuracy is significantly improved, with decreases of 19.9% in HD95 and 25.0% in ASSD—metrics that are particularly relevant for clinical tasks requiring precise anatomical delineation. The model’s architectural components—MRAB with directionally dilated convolutions and ASCAB —are designed to capture the anisotropic features of spinal structures while maintaining computational efficiency. Compared to the 3D U-Net, MRAS-FPN reduces training time by 23% and memory usage by 15% while delivering higher segmentation accuracy, making it suitable for clinical integration. Ablation studies confirm the importance of each module, with MRAB contributing a 3.9% increase in DSC and ASCAB adding a further 2.1%. MRAS-FPN maintains consistent performance across all vertebral levels (T9–S1) and intervertebral discs, supported by stable training dynamics and minimal signs of overfitting. Its ability to align technical performance with practical demands positions it as a strong candidate for integration into diagnostic workflows. This model recommends a clinically viable automated spine analysis tool and establishes a novel performance benchmark, reinforcing the role of intelligent automation in advancing radiological practice. Author contributions Conceptualisation, Methodology, Formal analysis, Investigation, Writing—Original Draft, Writing—Review & Editing, Visualization: Jeevika Tharini Viswanathan; Software: Vijaya Bhaskar Sadu and Beluguri Venkateswarlu; Validation, Writing—Review & Editing: Bharathi Subramaniam; Resources: Rukmani Devi Sethuraman; Data Curation: Sudhakar Tummala. Data availability Availability of Data and Materials: The MRSpineSeg Challenge dataset used in this study was originally released as part of the 2nd China Society of Image and Graphics (CSIG) Image and Graphics Technology Challenge: MRSpineSeg Challenge—Automated Multi-class Segmentation of Spinal Structures on Volumetric MR Images. The original challenge repository (https://www.spinesegmentation-challenge.com/) is no longer publicly accessible, and no verified active download link was available at the time of manuscript preparation. The dataset is therefore not redistributed with this study. Researchers seeking access to MRSpineSeg are advised to contact the original challenge organizers or the corresponding authors of published studies that have used this dataset.Code Availability: The source code and implementation details of the proposed framework are publicly available on GitHub at https://github.com/jeevikagv/Lumbar_MRI_Segmentation. Additionally, the archived version of the codebase and associated research materials are accessible via Zenodo at 10.5281/zenodo.20398840. Code availability The source code and implementation details of the proposed framework are publicly available on GitHub at https://github.com/jeevikagv/Lumbar_MRI_Segmentation. Additionally, the archived version of the codebase and associated research materials are accessible via Zenodo at https://doi.org/10.5281/zenodo.20398840. Declarations Competing interests The authors declare no competing interests. Ethical approval A statement to confirm that all methods were carried out in accordance with relevant guidelines and regulations.This study used a publicly available benchmark dataset (MRSpineSeg Challenge). Since the dataset was previously anonymized and released for research purposes, no additional ethical approval or informed consent was required for this retrospective computational study. Informed consent was obtained from all subjects and/or their legal guardian(s). Informed consent was obtained from all the experts. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Availability of Data and Materials: The MRSpineSeg Challenge dataset used in this study was originally released as part of the 2nd China Society of Image and Graphics (CSIG) Image and Graphics Technology Challenge: MRSpineSeg Challenge—Automated Multi-class Segmentation of Spinal Structures on Volumetric MR Images. The original challenge repository (https://www.spinesegmentation-challenge.com/) is no longer publicly accessible, and no verified active download link was available at the time of manuscript preparation. The dataset is therefore not redistributed with this study. Researchers seeking access to MRSpineSeg are advised to contact the original challenge organizers or the corresponding authors of published studies that have used this dataset.Code Availability: The source code and implementation details of the proposed framework are publicly available on GitHub at https://github.com/jeevikagv/Lumbar_MRI_Segmentation. Additionally, the archived version of the codebase and associated research materials are accessible via Zenodo at 10.5281/zenodo.20398840. The source code and implementation details of the proposed framework are publicly available on GitHub at https://github.com/jeevikagv/Lumbar_MRI_Segmentation. Additionally, the archived version of the codebase and associated research materials are accessible via Zenodo at https://doi.org/10.5281/zenodo.20398840.