Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation
Farhana Yasmin 0000-0002-5247-0958, Mahade Hasan 0009-0006-2778-1216, Haipeng Liu 0000-0002-4212-2503, Amjad Ali 0000-0002-5346-1017, Ghulam Muhammad 0000-0002-9781-3969, Yu Xue 0000-0002-9069-7547 · 2026
Read Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation by Farhana Yasmin 0000-0002-5247-0958, Mahade Hasan 0009-0006-2778-1216, Haipeng Liu 0000-0002-4212-2503, Amjad Ali 0000-0002-5346-1017, Ghulam Muhammad 0000-0002-9781-3969, Yu Xue 0000-0002-9069-7547 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-08238.txt.
Introduction
Farhana Yasmin, Mahade Hasan, Haipeng Liu, Amjad Ali, and coauthors search automatically for heart MRI segmentation networks that respect a compute budget. Outlining the ventricles is central to measuring cardiac function, but low contrast, fuzzy borders, and scan differences make it hard, and hand-designed networks may waste resources.
Their CardiacNAS framework evolves architectures from a UNet-like supernet over a cardiac-aware search space covering depth, width, kernels, attention, fusion, dropout, and residual scaling. Candidates are scored jointly on overlap accuracy and boundary distance against model size and floating point cost. The search runs under fixed budgets and is evaluated on the ACDC dataset against six leading methods.
Catalog record
| CREATOR | Farhana Yasmin 0000-0002-5247-0958, Mahade Hasan 0009-0006-2778-1216, Haipeng Liu 0000-0002-4212-2503, Amjad Ali 0000-0002-5346-1017, Ghulam Muhammad 0000-0002-9781-3969, Yu Xue 0000-0002-9069-7547 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08238
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation Farhana Yasmin 0000-0002-5247-0958, Mahade Hasan 0009-0006-2778-1216, Haipeng Liu 0000-0002-4212-2503, Amjad Ali 0000-0002-5346-1017, Ghulam Muhammad 0000-0002-9781-3969, Yu Xue 0000-0002-9069-7547 Abstract Cardiac magnetic resonance (CMR) segmentation underpins quantitative assessment of ventricular structure and function, yet reliable delineation remains difficult due to low tissue contrast, fuzzy boundaries, and inter-scan variability. We present CardiacNAS, an evolutionary neural architecture search (NAS) framework that couples a U-Net–like supernet with a cardiac-aware search space spanning depth/width, kernel size, filter size, attention, fusion, activation, dropout, and residual scaling. The search is explicitly resource-aware, jointly optimizing dice similarity coefficient (DSC) and 95th-percentile Hausdorff distance (HD95) versus model size and floating-point operations (FLOPs) under fixed compute budgets. Candidate architectures are instantiated from the supernet, trained with proxy budgets, and evolved through crossover, mutation, and elitist selection. We evaluate on the ACDC dataset and compare against six state-of-the-art methods, using qualitative comparisons, learning-curve analyses, and design-factor correlation studies. The resulting model attains 93.22% average DSC and 4.73,mm HD95 with 3.58M parameters and 14.56 GFLOPs, demonstrating a favorable accuracy-efficiency trade-off. Analyses indicate that searched attention and fusion choices, together with residual scaling, contribute to improved boundary fidelity and stability. CardiacNAS offers a principled, resource-aware approach to deployable CMR segmentation with transparent reporting of architectural complexity and compute budgets. I Introduction Cardiac magnetic resonance (CMR) segmentation enables quantitative assessment of ventricular anatomy and function by delineating the left ventricle (LV), right ventricle (RV), and myocardium (MYO) [ 1 ] [ 16 ] . Reliable contours are challenging due to low tissue contrast, fuzzy boundaries around papillary muscles and insertion points, inter-patient variability, and domain shift across scanners and protocols [ 10 ] [ 4 ] . These factors make boundary accuracy and efficiency central to downstream clinical use, where dice similarity coefficient (DSC) and the 95th-percentile Hausdorff distance (HD95) are standard indicators of region overlap and contour fidelity [ 14 ] [ 21 ] . Deep learning has become the dominant paradigm for medical image segmentation, with fully convolutional and encoder–decoder architectures forming the backbone [ 17 ] [ 6 ] [ 3 ] [ 20 ] . Successive designs add multi-scale context, denser skip connections, attention, and residual pathways [ 9 ] [ 27 ] [ 15 ] [ 7 ] . Despite progress, practitioners often face trade-offs between boundary quality and computational efficiency, which complicates deployment on resource-limited systems. Neural architecture search (NAS) aims to automate design under such constraints, spanning differentiable, evolutionary, and hybrid strategies [ 22 ] [ 23 ] [ 25 ] [ 8 ] . However, many pipelines still require substantial engineering to balance accuracy, FLOPs, and parameter counts. Fig. 1: Overall framework of CardiacNAS, showing supernet initialization, candidate generation through crossover and mutation, and evolutionary selection steps leading to the final architecture. To address these limitations, we present CardiacNAS , an evolutionary NAS framework designed for CMR segmentation while considering both accuracy and efficiency. The key contributions are as follows: • We propose a cardiac-aware supernet and search space that integrate multi-scale context, attention, and flexible skip-fusion to target crisp LV/MYO/RV boundaries under tight complexity budgets. • We propose an evolutionary NAS procedure with proxy training and resource-aware selection that explicitly trades off segmentation quality (DSC/HD95) against model complexity (parameters/FLOPs). • We provide a comprehensive evaluation, including qualitative comparisons, learning-curve analyses, and design-factor correlation studies that clarify how attention, fusion, and residual scaling influence boundary fidelity and stability. The rest of the paper is structured as follows: Section II reviews medical image segmentation and NAS methods; Section III details the CardiacNAS framework; Section IV reports the experimental setup, results, and analyses; and Section V summarizes the findings and outlines future work. II Related Works This section summarizes prior work on medical image segmentation and NAS-based segmentation frameworks. II-A Medical Image Segmentation Medical image segmentation enables quantitative analysis by delineating organs and lesions across MRI, CT, ultrasound, and endoscopy. Encoder–decoder CNNs remain central, beginning with U-Net [ 18 ] and extending to designs that enrich multi-scale context and efficiency including FTransDeepLab [ 5 ] and the self-configuring nnU-Net [ 11 ] . Hybrid CNN–Transformer models such as TransUNet [ 2 ] capture long-range dependencies while preserving local inductive biases. In contrast to hand-crafted backbones and single-objective setups, CardiacNAS couples a U-Net–like supernet with a cardiac-aware search space (depth/width, kernel size and dilation, attention, fusion, activation, dropout, residual scaling) and uses evolutionary search under explicit accuracy–efficiency constraints (DSC/HD95 vs. parameters/FLOPs), aligning with trends toward multi-objective, resource-aware design. II-B NAS-Based Segmentation Approaches NAS reduces manual design effort and systematically explores accuracy–efficiency trade-offs. Representative strategies include multi-objective cell/backbone search, such as MOE-FewSeg [ 12 ] and evolutionary search over compact operator/topology spaces, such as BiX-NAS [ 19 ] . Recent medical NAS frameworks expand to hybrid CNN–Transformer operators with efficient (single-/multi-path) supernets, including HCT-Net [ 24 ] and M-ENAS [ 13 ] . In contrast, CardiacNAS adopts a resource-aware NAS formulation balancing segmentation fidelity (DSC/HD95) against model complexity (Params/FLOPs) under explicit compute budgets and documented learning dynamics, enabling deployable cardiac segmentation. III Methods The proposed CardiacNAS framework integrates a supernet architecture with an evolutionary neural architecture search to design efficient models for cardiac MRI segmentation. As shown in Fig. 1 , the workflow begins with defining a supernet and a parameterized search space that controls depth, filter size, kernel size, attention modules, and fusion strategies. Candidate architectures are first generated randomly and then progressively improved through evolutionary operations such as crossover and mutation. Each candidate is trained with a limited budget and evaluated on segmentation accuracy and efficiency measures, while constraint checks ensure feasibility. Through repeated selection and elitism across generations, CardiacNAS converges toward high-performing architectures, from which the final model is chosen. III-A Supernet Architecture We employ a U-Net like supernet as the backbone of CardiacNAS, structured into an encoder, bottleneck, and decoder with skip connections. This framework provides the search space within which candidate architectures are instantiated and evolved. III-A 1 Input Each cardiac MR slice is denoted as x\in\mathbb{R}^{H\times W\times 1} , normalized prior to training. The network processes fixed-size inputs ( 128\times 128 ). TABLE I: Search Space Parameters Used for Architecture Generation in CardiacNAS Variable Search Space Filter sizes ( f_{l} ) [32 - 127] Kernel size ( k_{n} ) [1 - 7] Number of layers ( n_{s} ) [2 - 4] Dropout rate ( d_{p} ) [0.1 - 0.5] Attention modules ( a_{m} ) [squeeze-excitation, self-attention] Fusion/aggregation ( f_{a} ) [add, concat, weighted sum] Activation functions ( a_{c} ) [relu, elu, tanh, sigmoid] Residual scaling factor ( \alpha ) continuous [0.1 – 1.0] Algorithm 1 CardiacNAS: Neural Architecture Search 1: Input: training set \mathcal{D}_{tr} , validation set \mathcal{D}_{va} , search space \Omega , population size M , generations T , crossover rate \pi_{\chi} , mutation rate \pi_{\mu} 2: Output: best architecture \hat{\gamma} 3: // Initialization 4: Sample initial population \mathcal{G}^{(0)}=\{\gamma_{1},\ldots,\gamma_{M}\} with \gamma_{i}\sim\Omega (genotypes over variables in Table I ) 5: \hat{\gamma}\leftarrow\varnothing , \hat{f}\leftarrow-\infty 6: for t=1 to T do 7: // Parent selection (e.g., tournament) 8: Build mating pool \mathcal{M} by repeatedly selecting parents from \mathcal{G}^{(t-1)} 9: // Variation: crossover & mutation 10: \mathcal{O}\leftarrow\emptyset (offspring) 11: while |\mathcal{O}|<M do 12: Pick (\gamma^{a},\gamma^{b}) from \mathcal{M} 13: