The diagnostic value of MRI segmentation technique for shoulder joint injuries based on deep learning Authors: Lina Dai, Md Gapar Md Johar, Mohammed Hazim Alkawaz Published: 2024 ABSTRACT AbstractThis work is to investigate the diagnostic value of a deep learning-based magnetic resonance imaging (MRI) image segmentation (IS) technique for shoulder joint injuries (SJIs) in swimmers. A novel multi-scale feature fusion network (MSFFN) is developed by optimizing and integrating the AlexNet and U-Net algorithms for the segmentation of MRI images of the shoulder joint. The model is evaluated using metrics such as the Dice similarity coefficient (DSC), positive predictive value (PPV), and sensitivity (SE). A cohort of 52 swimmers with SJIs from Guangzhou Hospital serve as the subjects for this study, wherein the accuracy of the developed shoulder joint MRI IS model in diagnosing swimmers’ SJIs is analyzed. The results reveal that the DSC for segmenting joint bones in MRI images based on the MSFFN algorithm is 92.65%, with PPV of 95.83% and SE of 96.30%. Similarly, the DSC for segmenting humerus bones in MRI images is 92.93%, with PPV of 95.56% and SE of 92.78%. The MRI IS algorithm exhibits an accuracy of 86.54% in diagnosing types of SJIs in swimmers, surpassing the conventional diagnostic accuracy of 71.15%. The consistency between the diagnostic results of complete tear, superior surface tear, inferior surface tear, and intratendinous tear of SJIs in swimmers and arthroscopic diagnostic results yield a Kappa value of 0.785 and an accuracy of 87.89%. These findings underscore the significant diagnostic value and potential of the MRI IS technique based on the MSFFN algorithm in diagnosing SJIs in swimmers.Keywords: Deep learning, MSFFN algorithm, MRI image segmentation, Swimmers, Shoulder joint injuries, DiagnosisSubject terms: Mathematics and computing, Applied mathematics, Computational science, Computer science, Information technology, Pure mathematics, Scientific data, Software, Statistics Introduction Shoulder joint injury (SJI) is one of the common sports injuries among swimmers, which may significantly impact their training and competition and potentially affect their quality of life1. Magnetic resonance imaging (MRI) technology is applied for diagnosing SJIs due to its high resolution and multidimensional data capabilities. However, traditional MRI image analysis methods often rely on subjective visual assessments by physicians, leading to uncertainty and variability in diagnostic outcomes2. Deep learning (DL)-based MRI image segmentation (IS) techniques offer a novel approach to address this challenge by automatically identifying and segmenting different tissues and structures within MRI images, thus providing clinicians with more accurate and objective diagnostic information3,4. Therefore, the adoption and diagnostic value of DL-based MRI IS techniques in diagnosing SJIs in swimmers are assessed, aiming to provide more reliable auxiliary diagnostic tools for clinical practice and more effective support for swimmers’ health and rehabilitation. Research background and motivations SJIs are common issues throughout athletes’ careers, particularly in sports that require frequent use of shoulder strength. Swimming, being a high-intensity activity, demands frequent shoulder movements, posing challenges to the stability and function of the shoulder joint5. These injuries not only affect athletes’ performance and careers but may also result in long-term functional impairments and pain. Therefore, timely and accurate diagnosis and treatment of SJIs are crucial for the health and professional development of swimmers. Currently, clinical diagnosis of SJIs involves imaging techniques and arthroscopic examinations. Among these methods, MRI provides high-resolution images, offering clear visualization of soft tissue structures. Moreover, functional information about the shoulder joint can be obtained through techniques such as dynamic MRI and magnetic resonance arthrography, aiding in a more comprehensive assessment of the extent and impact of injuries6. However, traditional approaches have certain limitations, including dependence on the experience and skill level of physicians, as well as issues related to the accuracy of diagnosing soft tissue injuries in the shoulder7. In recent years, with the advancement and adoption of DL technology, MRI IS techniques based on DL have gradually become a research hotspot. DL technology has made breakthrough progress in the field of medical imaging due to its excellent image processing and feature extraction capabilities8. Through DL algorithms, automated segmentation of MRI images can be achieved, extracting crucial anatomical structure information such as muscles, tendons, and ligaments. This aids physicians in more accurately diagnosing SJIs and assessing their extent and type9. DL-based MRI IS techniques have already achieved significant success in other fields such as tumor detection and brain disease diagnosis10,11, offering new approaches and methods for diagnosing SJIs. In the field of DL applications, fully convolutional neural networks (CNNs) have emerged as crucial tools in medical image analysis, particularly the U-Net and AlexNet models. U-Net is specifically designed for medical IS, employing a symmetrical encoder-decoder architecture to achieve precise segmentation. Conversely, AlexNet is a classic CNN that excels in image classification tasks. However, both models exhibit limitations; for instance, U-Net may struggle to fully leverage global contextual information when processing diverse and multi-scale images, while AlexNet’s relatively simple architecture lacks specific optimizations for medical image features, potentially leading to imprecise segmentation in images with intricate details or complex structures. Furthermore, research on DL techniques for shoulder injuries in swimmers is still in its infancy. Such injuries possess unique characteristics, including muscle fatigue and imbalance resulting from repetitive actions, which may impact the accuracy and stability of DL algorithms. Thus, there is a pressing need to improve U-Net and AlexNet models to develop a segmentation technique suitable for MRI images of shoulder injuries in swimmers. This exploration aims to uncover their potential in enhancing the accuracy and efficiency of diagnosing SJIs, providing new approaches and methods for ensuring the health of athletes. Research objectives A segmentation model is established for shoulder joint MRI images of swimmers based on the fully CNNs U-Net and AlexNet. Through experimentation and data analysis, the accuracy, stability, and reliability of the established segmentation model in diagnosing SJIs in swimmers are evaluated and compared with traditional methods. The potential of DL-based MRI IS techniques in enhancing the accuracy and efficiency of diagnosing SJIs in swimmers is explored, aiming to provide new diagnostic methods and technological support for ensuring the health of athletes and promoting their health and professional development. Literature review The shoulder joint, constituted by the articulation between the humeral head and the glenoid cavity of the scapula, is one of the most mobile joints in the human body. Repetitive swimming strokes can lead to laxity in the anterior shoulder joint capsule, being one of the primary causes of swimmer’s shoulder12. During abduction of the shoulder joint, blood vessels become engorged with blood; however, inadequate blood supply occurs during internal rotation and adduction, resulting in the extrusion of blood from the tendons, leading to the accumulation of waste products, thereby predisposing to muscle soreness, stiffness, strains, and aseptic inflammation over time13. Hence, early diagnosis and treatment hold significant importance. MRI IS is crucial in diagnosis of SJIs. It accurately locates soft tissue damage in the shoulder joint, facilitating the early detection of potential injuries, including soft tissue inflammation, tumors, and other lesions14. This aids in early intervention and treatment, thereby preventing disease progression. Traditional MRI bone region segmentation methods, including thresholding, region growing, and level set methods, have been applied in shoulder joint segmentation. However, these methods often require prior knowledge, lack complete automation for accurate detection and segmentation, and exhibit relatively low universality and precision, thus having significant limitations15,16. Some researchers propose using pixel-centered image blocks as training and prediction samples for CNNs to achieve pixel-level classification. However, this method suffers from the problem of strong local perceptual ability but insufficient global feature perception17. To address this issue, researchers employed fully convolutional networks (FCNs) for semantic segmentation of images. Currently, there is limited research on the adoption of DL algorithms in bone joint segmentation. Memiş et al.18 utilized deep CNNs for two-dimensional semantic segmentation of the femoral head and proximal femur in MRI images of patients with arthritis, obtaining 194 low-quality MRI slices from 33 MRI sequences of 13 patients. More recently, Cantarelli Rodrigues et al.19 developed a fully automated segmentation model using CNNs for proton density-weighted images, and applied it to shoulder joint MRI IS. They achieved high accuracy in both humeral and glenoid joint segmentation. Kaniewska et al.20 revealed that accelerated PROPELLER sequences with DL post-processing exhibited superior image quality and higher diagnostic confidence relative to traditional sequences. DL sequences allowed for comprehensive evaluation of the supraspinatus tendon and yielded consistent results with traditional sequences in assessing other structures of the shoulder joint. Compared to FCNs, the U-Net model demonstrates superior accuracy in biomedical IS, primarily due to its design incorporating skip connections, which enhances its ability to segment complex structures and detailed images, especially with small sample sizes. U-Net al.so excels in bone segmentation, as seen in MRI images of joints where the humeral head exhibits significant variability, while other regions share similar shapes and gray-level characteristics. Block-based local segmentation methods may offer higher accuracy in delineating local boundary contours. However, the complexity of the U-Net architecture results in substantial computational demands, particularly when processing high-resolution images, which can lead to prolonged training times and increased memory usage. Therefore, this study proposes an improved and integrated approach combining AlexNet and U-Net algorithms to create a novel multi-scale feature fusion network (MSFFN) for the detection and segmentation of the region of interest corresponding to the humeral head. MSFFN model constraution In the MSFFN model, the optimized AlexNet and U-Net are integrated. This work employs the improved Alex-Net algorithm to address two aspects of shoulder joint segmentation. Firstly, it segments the regions of interest (ROIs) in shoulder joint injury MRI images based on the Alex-Net algorithm to generate a segmentation model. Subsequently, the improved U-Net algorithm is utilized to further detect and finely segment the bone regions within the images. Specifically, the U-Net algorithm segments the joint bones and humerus separately within the ROIs of MRI images. Then, false bone regions are filtered based on their positions and sizes, and a voting filter is applied. Additionally, inter-frame prediction is employed to address missing bone and muscle regions, ultimately yielding finely segmented results of the bone ROI. The specific workflow for MRI IS of SJIs in this work is illustrated in Fig. 1. MRI IS process for shoulder joint injury. Segmentation of shoulder joint roi based on ALEXNET algorithm The AlexNet architecture comprises eight layers and demonstrates remarkable performance in image recognition21. MRI image blocks are square images centered around pixels, with pixels selected within a range of 5–10 pixels from the nearest bone edge. The traditional AlexNet al.gorithm is primarily designed for processing red, green, blue (RGB) input images, whereas MRI images are grayscale. Hence, three Window-Level mappings or sampling the image blocks at three different resolutions are required to increase them to three channels. However, this approach does not improve performance compared to using a single channel for testing22, thus the input channels are adjusted to one. The MRI input images processed by AlexNet have a resolution of 224 × 224 pixels, while the resolution range of the image blocks is 7 × 7 pixels to 23 × 23 pixels. Consequently, adjustments are made to the AlexNet al.gorithm structure as follows: (1) adjusting the input channels to one; (2) setting the convolutional layer filter size to 3 × 3 pixels with a stride of 1; (3) adjusting the pool layer kernel size to 2 × 2 pixels; (4) modifying the feature channels to 64, 64, 128, 128, and 256. The modified AlexNet al.gorithm structure is illustrated in Fig. 2. AlexNet algorithm structure after integration. The AlexNet model performs segmentation on binary images, utilizing a 3 × 3 pixel kernel for two iterations of erosion and two iterations of dilation to reduce noise and connect bone regions. The erosion operation is described by Eq. (1): where I represents the input image, I denotes the structuring element (i.e., the 3 × 3 pixel kernel), and I′ is the eroded image. where I′ is the eroded image and I′′ is the dilated image. Subsequently, position and size filtering is applied within the candidate bone regions to remove boundary locations and excessively large or small rectangles, as outlined in Eq. (3): where Ri is the candidate region, area (Ri) denotes the area of the region, and Amin and Amax represent the minimum and maximum allowed area, respectively. To encompass all ROI while minimizing the introduction of excess noise, the boundaries of the bone regions are expanded outward by ten pixels, as described in Eq. (4): where R is the original bone region and R extended is the expanded bone region. The longest closed contour is then extracted from the expanded bone rectangular region to obtain precise edges of the bone. This process is achieved through a contour detection algorithm (Canny edge detection), as illustrated in Eq. (5): where C represents the extracted contour, utilizing an appropriate edge detection algorithm to achieve accurate edges. Segmentation of shoulder joint roi based on improved U-NET algorithm The lightweight U-Net is a network specifically designed for biomedical IS tasks, utilizing an encoder-decoder model structure (Fig. 3). While U-Net has demonstrated remarkable performance in the field of medical IS, its complex architecture results in high computational demands and significant memory requirements. Additionally, U-Net may not fully exploit multi-scale information when processing images with features of varying scales. Furthermore, standard convolutions are often inadequate for accommodating the feature distributions of different images, leading to suboptimal feature extraction. Therefore, the following improvements are proposed for the lightweight U-Net: U-Net algorithm structure diagram. Design of the encoding stage In the encoding stage, the network extracts feature information from the input image through a series of convolutional modules, employing max pooling operations to reduce resolution. This approach expands the receptive field of the convolutional operations, thereby capturing more global information. The stage consists of four small modules, each comprising two cascaded 3 × 3 convolution operations, which aim to maintain the same receptive field size as a 5 × 5 convolution kernel while significantly reducing the number of parameters. Specifically, the number of channels is doubled in each small module to extract more high-level semantic features. This design not only reduces computational complexity but also ensures better performance during feature extraction. Design of the decoding stage The decoding stage employs a design that is symmetric to the encoding stage, replacing max pooling operations with transposed convolution operations to restore the resolution of the feature maps. In each small module, the number of channels is halved, effectively utilizing the features extracted during the encoding phase and alleviating the computational burden of the model. Enhancement of the resolution of input images To maintain the resolution of the input MRI images, the study employs mirroring operations to enhance the images, thereby improving pixel resolution. This method effectively mitigates potential information loss during convolution operations, resulting in increased accuracy in the final segmentation. Improved feature fusion and adaptive convolution To address the model’s inability to adapt to different feature distributions in images, a multi-scale feature fusion approach is utilized, introducing features of varying scales during the decoding stage to enhance the model’s capability to perceive multi-scale information. Let F1, F2, F3 represent the feature maps of different scales in the decoding stage. These feature maps are aligned through upsampling before concatenation, as shown in Eq. (6). This is followed by the introduction of learned fusion weights α1, α2, α3, as represented in Eq. (7). To address the issue of adapting to different feature distributions in images, the study employs an Adaptive Convolution strategy that dynamically adjusts the convolution kernels based on the features of the input images. A small neural network N is utilized to generate the adaptive convolution kernel K, as illustrated in Eq. (8): where X represents the input feature map. The adaptive convolution is then applied to the input feature map F, as shown in Eq. (9): where ∗ denotes the convolution operation. Experimental design and performance evaluation Experimental materials The dataset comprises MRI images of SJIs diagnosed from Guangzhou Hospital, consisting of 5 groups of scans from swimmers, with approximately 2000 images per scanning sequence. The details of MRI imaging and annotations are provided in Table 1. Among these, 3 groups of patients are utilized for training and undergoing 5-fold cross-validation, while the remaining 2 groups are reserved for testing segmentation effectiveness. All MRI datasets are converted into 8-bit grayscale images with a resolution of 192 × 192 pixels and are denoised using Gaussian filtering. Medical experts annotate the joint bones and humeral head regions, and the dataset is partitioned into training, validation, and test sets to support evaluation of DL models. To further validate the effectiveness of the model, the study conducts external validation using MRI data from external research institutions for testing, confirming the model’s efficacy. Table 1Details of MRI imaging and annotations.ParameterDescriptionImaging sequencesT1-weighted (T1W), T2-weighted (T2W), Proton density (PD)Imaging planesAxial, coronal, sagittalScanner modelSiemens Magnetom Skyra 3.0Tscanning Parameters—Magnetic field strength: 3.0T—Slice thickness: 3 mm—Slice spacing: 0.5 mm—Matrix size: 512 × 512—Field of view (FOV): 250 mm x 250 mmReconstruction algorithmFast adaptive reconstructionImaging timeT1W: 5 min, T2W: 6 min, PD: 5 minSignal intensity (SNR)30–50 dBAnnotation methodManual annotation of joint bones and humeral head regions by experienced radiologists Details of MRI imaging and annotations. The work includes 52 swimmer patients with SJIs from Guangzhou Hospital, all of whom underwent MRI examination. Among them, there were 29 male and 23 female patients. There were 42 cases of acute injuries and 10 cases of chronic injuries. The patients were 18 –35 years old, with an average duration of swimming activity of (8.03 ± 3.11) years and heights ranging from 168 to 190 cm. Inclusion criteria: (1) Age over 18 years old; (2) Shoulder joint injury or restricted movement caused by swimming activities; (3) First-time shoulder injury; (4) Patients with unilateral shoulder joint damage. Exclusion criteria: (1) Patients simultaneously suffering from wrist joint injuries or other joint injuries; (2) Patients requiring emergency surgery; (3) Patients who did not undergo MRI examination. This work was approved by the Guangzhou College of Commerce Ethics Committee. The subjects were informed about and agreed to participate in this work and signed informed consent forms. Experimental environment The experimental setup consists of a computer with a CPU E5-2678 v3, 128GB of memory, and 4 NVIDIA GeForce RTX 2080Ti graphics cards. All algorithms are implemented under the PyTorch framework, with a learning rate of 0.01, a learning rate decay coefficient of 0.5, weight decay of 0.0001, momentum coefficient of 0.9, batch size of 128, dropout parameter of 0.5, and 150,000 training iterations. To enhance training efficiency, synchronized batch normalization (BN) operations across GPUs are utilized with a batch size of 20, and the preset number of training epochs is 50. This configuration and parameter selection support the training of complex models, and the optimization and synchronized BN operations enhance training speed and effectiveness. Parameters setting For training the MSFFN model, data augmentation is performed on the original training set by applying methods such as random horizontal flipping, random vertical flipping, random shifts within a 10% range of height and width, and random 90° rotations. Batch size, number of iterations, and epochs are set, with corresponding values of momentum set to 0.2, batch size of 1, 128 iterations, and 200 epochs. Performance evaluation Different sizes of MRI image blocks are trained to validate their performance in MRI IS (Fig. 4). It is observed that when the image block size is 11 × 11, the segmentation accuracy reaches its highest value of 98%. MRI IS accuracy with different image block sizes. The dice similarity coefficient (DSC), positive predictive value (PPV), volume overlap error (VOE), relative volume difference (RVD), and sensitivity (SE) reflect the segmentation outcomes. The equations for calculating DSC, PPV, and SE are as follows: In the equations, A represents the segmented voxel region by the algorithm model, and B represents the segmented voxel region by manual segmentation. The study evaluates the segmentation performance of the MSFFN model, AlexNet, U-Net, Statistical Shape Model (SSM)23, and CNN segmentation models24, using metrics including DSC, PPV, SE, VOE, and RVD for comparison. The DSC results are presented in Fig. 5. The MSFFN model achieves a DSC of (92.65 ± 1.21)% for joint bone segmentation in MRI images, compared to (85.42 ± 2.03)% for AlexNet, (86.31 ± 1.72)% for U-Net, (90.33 ± 4.21)% for the SSM algorithm, and (89.71 ± 3.09)% for the CNN algorithm. For humeral segmentation, the MSFFN model attains a DSC of (92.93 ± 2.11)%, while AlexNet achieves (83.78 ± 1.34)%, U-Net (84.21 ± 2.17)%, SSM (91.54 ± 3.44)%, and CNN (88.65 ± 4.82)%. Statistical analysis demonstrates that the MSFFN model significantly outperforms the other models in DSC results for both joint bone and humeral segmentation, with AlexNet and U-Net showing notably lower values (P < 0.05). Specifically, the DSC of the MSFFN model exceeds that of AlexNet by 7.23%, U-Net by 6.34%, SSM by 2.32%, and CNN by 2.94% for joint bone segmentation. For humeral segmentation, the MSFFN model surpasses AlexNet by 9.15%, U-Net by 8.72%, SSM by 1.39%, and CNN by 4.28%. Comparison of DSC values for different algorithms (A: shoulder joint; B: humerus; “*”: compared to AlexNet, U-Net, SSM, and CNN models, P < 0.05; “#”: compared to SSM and CNN models, P < 0.05.) The PPV results are presented in Fig. 6. The MSFFN model achieves a PPV of (95.83 ± 2.11)% for joint bone segmentation in MRI images, compared to (86.45 ± 2.45)% for AlexNet, (88.12 ± 1.78)% for U-Net, (94.86 ± 2.44)% for the Statistical Shape Model (SSM), and (93.35 ± 3.45)% for the CNN algorithm. For humeral segmentation, the MSFFN model reports a PPV of (95.56 ± 2.78)%, while AlexNet, U-Net, SSM, and CNN yield PPV values of (87.21 ± 2.19)%, (88.38 ± 2.02)%, (94.71 ± 3.24)%, and (93.40 ± 3.20)%, respectively. Statistical analysis indicates that the MSFFN model significantly outperforms the other models in PPV for both joint bone and humeral segmentation, with AlexNet and U-Net exhibiting notably lower values (P < 0.05). Specifically, in joint bone segmentation, the PPV of the MSFFN model is higher by 9.38% compared to AlexNet, 7.71% compared to U-Net, 0.97% compared to SSM, and 2.48% compared to the CNN algorithm. In humeral segmentation, the MSFFN model also demonstrates superior performance, exceeding AlexNet by 8.35%, U-Net by 7.18%, SSM by 0.85%, and CNN by 2.16%. Comparison of PPV values for different algorithms (A: shoulder joint; B: humerus; “*”: compared to AlexNet, U-Net, SSM, and CNN models, P < 0.05; “#”: compared to SSM and CNN models, P < 0.05.) The SE results are presented in Fig. 7. The MSFFN model achieves an SE of (96.30 ± 2.32)% for joint bone segmentation in MRI images, compared to (81.42 ± 2.12)% for AlexNet, (82.22 ± 1.34)% for U-Net, (93.29 ± 2.23)% for the SSM, and (90.45 ± 3.02)% for the CNN algorithm. For humeral segmentation, the MSFFN model reports an SE of (92.78 ± 2.73)%, while AlexNet, U-Net, SSM, and CNN yield SE values of (81.28 ± 2.09)%, (84.33 ± 2.34)%, (88.01 ± 3.09)%, and (85.72 ± 2.12)%, respectively. Statistical analysis indicates that the MSFFN model significantly outperforms the other models in SE for both joint bone and humeral segmentation, with AlexNet and U-Net exhibiting notably lower values (P < 0.05). Specifically, in joint bone segmentation, the SE of the MSFFN model is higher by 14.88% compared to AlexNet, 14.08% compared to U-Net, 3.01% compared to SSM, and 5.85% compared to the CNN algorithm. In humeral segmentation, the MSFFN model’s SE exceeds AlexNet by 11.50%, U-Net by 8.45%, SSM by 4.77%, and CNN by 7.06%. Comparison of SE values for different algorithms (A: shoulder joint; B: humerus; “*”: compared to AlexNet, U-Net, SSM, and CNN models, P < 0.05; “#”: compared to SSM and CNN models, P < 0.05.) The VOE results are shown in Fig. 8. The MSFFN model exhibits a VOE of (2.78 ± 0.78)% for joint bone segmentation in MRI images, compared to (3.21 ± 0.27)% for AlexNet, (3.17 ± 0.24)% for U-Net, (2.83 ± 0.28)% for the SSM, and (3.02 ± 0.12)% for the CNN algorithm. For humeral segmentation, the MSFFN model achieves a VOE of (2.83 ± 0.73)%, while AlexNet, U-Net, SSM, and CNN report VOEs of (3.32 ± 0.19)%, (3.03 ± 0.26)%, (2.90 ± 0.09)%, and (2.96 ± 1.12)%, respectively. Statistical analysis reveals that the MSFFN model significantly outperforms the other models in VOE for both joint bone and humeral segmentation, with AlexNet and U-Net showing notably higher values (P < 0.05). Specifically, for joint bone segmentation, the VOE of the MSFFN model is lower by 0.43% points compared to AlexNet, 0.39% points compared to U-Net, 0.05% points compared to SSM, and 0.24% points compared to the CNN algorithm. In the case of humeral segmentation, the VOE is reduced by 0.49% points compared to AlexNet, 0.20% points compared to U-Net, 0.07% points compared to SSM, and 0.13% points compared to the CNN algorithm. Comparison of VOE values for different algorithms (A: shoulder joint; B: humerus; “*”: compared to AlexNet, U-Net, SSM, and CNN models, P < 0.05; “#”: compared to SSM and CNN models, P < 0.05.) The RVD results are presented in Fig. 9. Notably, the MSFFN model achieves an RVD of (0.60 ± 0.08)% for joint bone segmentation in MRI images, compared to (0.78 ± 0.07)% for AlexNet, (0.74 ± 0.08)% for U-Net, (0.65 ± 0.06)% for the SSM, and (0.71 ± 0.07)% for the CNN algorithm. For humeral segmentation, the MSFFN model reports an RVD of (0.57 ± 0.73)%, while AlexNet, U-Net, SSM, and CNN have RVDs of (0.72 ± 0.11)%, (0.70 ± 0.16)%, (0.61 ± 0.09)%, and (0.67 ± 0.18)%, respectively. Statistical analysis indicates that the MSFFN model significantly outperforms the other models in RVD for both joint bone and humeral segmentation (P < 0.05). Specifically, the RVD for joint bone segmentation is reduced by 0.18% compared to AlexNet, 0.14% compared to U-Net, 0.05% compared to SSM, and 0.11% compared to the CNN algorithm. For humeral segmentation, the RVD is reduced by 0.15% compared to AlexNet, 0.13% compared to U-Net, 0.04% compared to SSM, and 0.10% compared to the CNN algorithm. Comparison of RVD values for different algorithms (A: shoulder joint; B: humerus; “*”: compared to AlexNet, U-Net, SSM, and CNN models, P < 0.05; “#”: compared to SSM and CNN models, P < 0.05.) Using both conventional MRI and the segmentation model developed in this work, the diagnosis of SJIs in swimmers is conducted. The diagnostic results of shoulder joint injury types in swimmers are presented in Table 2. The accuracy of diagnosing shoulder joint injury types using the MRI IS algorithm developed in this work is 86.54%, which is higher than the conventional diagnostic accuracy of 71.15%. Table 2Accuracy of MRI diagnosis of shoulder joint injury types in swimmers.Damage typeRoutine MRIThe MRI processed by the model in this workShoulder sleeve injury (30)23 (76.67)27 (90.00)Shoulder joint tear (15)10 (66.67)12 (80.00)Joint instability (7)4 (57.14)6 (85.71)Total (52)37 (71.15)45 (86.54) Accuracy of MRI diagnosis of shoulder joint injury types in swimmers. The diagnostic results of shoulder joint injury severity obtained using the MRI IS model developed in this work are compared with those of the gold standard arthroscopy. The comparison reveals a consistency Kappa value of 0.785 and an accuracy of 87.89% between the diagnostic results of complete tear, partial tear on the superior surface, partial tear on the inferior surface, and intra-tendinous tear of the shoulder joint among swimmers obtained using the MRI IS model developed in this work and those obtained using arthroscopy. Conventional MRI and the MRI IS model developed in this work are respectively applied to process the images. Then, on the coronal images of the proximal humerus, the inclination angle (IA) and the distance from humeral head center to humeral tuberosity (THCD) are measured; on the axial images of the humerus, the retroversion angle of the humeral head (HTRA) and the version angle (VA) of the head are measured. The changes in the above indicators of the humeral head after processing by the two methods are compared. The results shown in Table 3 indicate that there are neglectable differences (P > 0.05) in IA, THCD, and VA detected on the affected side MRI images between conventional MRI and the MRI IS model developed in this work relative to the healthy side. However, the HTRA detected on the affected side MRI images after processing by both conventional MRI and the MRI IS model developed in this work are markedly lower than those detected (P < 0.05). Table 3Comparison of humeral head indicators.Damage typeRoutine MRIThe MRI processed by the model in this workIA (°)Affected side132.68 ± 5.03132.55 ± 4.65Healthy side133.11 ± 4.96132.87 ± 4.73THCD (cm)Affected side2.50 ± 0.182.55 ± 0.20Healthy side2.42 ± 0.192.46 ± 0.26HTRA (cm)Affected side24.22 ± 2.3424.67 ± 2.64Healthy side21.35 ± 2.49**20.56 ± 2.81**VA (°)Affected side22.78 ± 2.8522.54 ± 5.19Healthy side23.49 ± 2.7423.91 ± 2.65 Comparison of humeral head indicators. Discussion Due to the limited sample size included in the training process of our constructed algorithm, it is not feasible to compare it with MRI bone segmentation methods based on 3D FCNs as reported in previous studies25,26. Mu et al.27 established segmentation models based on DL algorithms to obtain candidate bone regions and detect positions of the humerus and joint bones through voting. The detected bone regions were further segmented using the AlexNet segmentation model to achieve pixel-level accuracy in bone edge segmentation. The results reveal that the model achieves mean accuracy of 0.91 ± 0.02, 0.95 ± 0.03, and 0.95 ± 0.02 for DSC, PPV, and SE, respectively. Kwolek et al.28 achieved an accuracy of 95.9% in segmenting the proximal tibia and patella from lateral X-ray images using U-Net. This indicates that both U-Net and AlexNet demonstrate strong performance in bone tissue segmentation. However, AlexNet, due to its relatively shallow architecture, tends to lose local detail and spatial information when processing high-resolution MRI data, adversely affecting segmentation accuracy. Although U-Net retains image features through skip connections, it often encounters issues with edge detail loss when dealing with irregular shapes and boundaries. This study proposes an improvement by optimizing and integrating U-Net with AlexNet to form the MSFFN model. The results indicate that the MSFFN model yields the highest values for DSC, PPV, and SE in segmenting joint bones and the humerus. The MSFFN model, through multi-scale feature fusion and effective retention of detailed information, demonstrates precise segmentation of complex bone structures and enhanced SE. These advantages primarily arise from the MSFFN model’s multi-scale design at the feature extraction level, which effectively merges feature information across different levels, thereby improving the model’s ability to capture objects of varying sizes and shapes29. Compared to traditional AlexNet and U-Net, the MSFFN model is better suited for comprehensive coverage of bone regions in complex anatomical environments, effectively mitigating the issue of incomplete segmentation, particularly in the segmentation of irregularly shaped humeral and joint bones. Furthermore, as important metrics for assessing segmentation accuracy, VOE and RVD demonstrate that the MSFFN model significantly outperforms other models. The MSFFN model achieves the lowest VOE value of 2.78% for joint bone segmentation, markedly lower than that of the other models, indicating its superior segmentation accuracy. The RVD values also reflect the MSFFN model’s advantage in reducing volumetric error, with both joint bone and humeral segmentations exhibiting lower RVDs compared to other models, signifying enhanced stability in volumetric consistency. VOE measures the overlap error between the segmentation results and the standard segmentation, while RVD evaluates the volume differences of the segmented regions30. The MSFFN model excels in both metrics, highlighting its superiority in segmentation precision and consistency. By leveraging multi-scale feature fusion and integrating contextual information, the MSFFN model effectively reduces deviations in the segmented areas, thereby improving both volumetric consistency and segmentation accuracy. This is particularly crucial for anatomical structures like joint bones and the humerus, where precise volume and boundary segmentation are essential for subsequent clinical diagnosis and treatment planning. In MRI IS, traditional segmentation methods such as threshold-based methods and edge detection methods31,32 are applicable for segmenting bone regions in various MRI images. A vast body of literature33–36 has established segmentation methods specifically tailored for knee joint segmentation. Traditional methods primarily rely on pixel grayscale features and existing boundary information, unable to learn inherent shape and positional of shoulder joint structures. Hence, these methods tend to exhibit lower accuracy across various metric parameters. Knee joints and shoulder joints are anatomically distinct structures with different distributions of bones and soft tissues. Hence, during IS, it is essential to consider the anatomical characteristics and morphological variations of different joints. Moreover, the types of common pathologies in knee and shoulder joints also differ, leading to variations in corresponding segmentation algorithms. For instance, common knee joint pathologies include osteoarthritis and ligament injuries, while shoulder joint pathologies often include rotator cuff injuries and frozen shoulder. Consequently, there exist differences between shoulder joint MRI IS and knee joint segmentation. In swimming, SJIs often occur due to incorrect arm movements, especially during rapid response situations. Currently, arthroscopic examination remains the gold standard in clinical practice, but its invasive nature imposes certain limitations on clinical adoptions. Therefore, actively seeking non-invasive and highly accurate examination methods has always been a focus of clinical research37. Currently, MRI is a commonly applied non-invasive technique, known for its ease of operation, high resolution, and clear imaging, making it widely applicable in diagnosing SJIs38. However, conventional MRI diagnosis of SJIs is not always absolutely accurate; sometimes, different types of injuries may exhibit similar signal characteristics, leading to difficulties in accurate differentiation and poor specificity. Additionally, MRI images may be influenced by factors such as structural overlap, poor transparency, or postoperative changes, causing some structures to be inadequately evaluated, such as the cartilage on the shoulder peak and tendon sheaths. This work applies the established MRI IS model to SJIs in swimmers. The results reveal that the MRI IS algorithm developed in this work achieves an accuracy of 86.54% in diagnosing SJIs in swimmers, which is higher than the conventional diagnostic accuracy of 71.15%. This indicates that the new algorithm proposed in the study achieved better diagnostic results in handling SJIs in swimmers. The segmentation model constructed in this work may better adapt to the characteristics of SJIs in swimmers, thereby improving segmentation accuracy. The segmentation technique used in this work may be more detailed and advanced, enabling better differentiation of different types of SJIs. Overall, the MRI IS algorithm developed in this work demonstrates higher accuracy in diagnosing SJIs in swimmers. Conclusion Research contribution This study proposes an MRI IS algorithm based on the MSFFN model, significantly enhancing the segmentation accuracy of joint bones and the humerus. Through multi-scale feature fusion and the integration of contextual information, the MSFFN model outperforms existing traditional models such as AlexNet and U-Net across multiple metrics, including DSC, PPV, SE, VOE, and RVD. It particularly demonstrates strong robustness and precision when handling irregular shapes and complex bone structures. Additionally, this algorithm is successfully applied to the MRI IS of SJIs in swimmers, achieving a diagnostic accuracy of 86.54%, representing a significant improvement over traditional methods. This result indicates that the MSFFN-based segmentation approach holds substantial potential for diagnosing SJIs in swimmers, providing reliable technical support for future non-invasive clinical diagnostics. Future works and research limitations Despite the significant advancements in segmentation accuracy and clinical application achieved in this study, certain limitations remain. This study only validates the model using our institution’s dataset. Due to challenges in data acquisition and issues related to data heterogeneity, external datasets were not used for validation. In future work, we will further validate and optimize the model using external datasets. Additionally, the relatively small sample size used in model training may affect the model’s generalization ability. Small datasets are prone to overfitting during training, leading to unstable performance in real-world applications. Future research should consider increasing the sample size, particularly through multi-center data collection, to enhance the model’s applicability across different clinical environments. Due to the limited sample size, direct comparisons with other 3D FCN-based MRI bone segmentation methods could not be made, leaving the performance of the MSFFN model in larger sample sizes and more complex datasets yet to be fully validated. Future work could incorporate more comparative analyses to comprehensively assess the superiority of MSFFN. Additionally, future research could explore multimodal segmentation approaches that combine other imaging techniques (such as CT or ultrasound) to further enhance segmentation capabilities for complex anatomical structures. Multimodal imaging can provide distinct tissue characteristics, aiding in improving segmentation accuracy and consistency to better meet clinical needs. Further studies may also focus on optimizing the architecture and algorithms of the MSFFN model to enhance its segmentation accuracy across different joints and pathological conditions. For example, the incorporation of advanced DL techniques, such as self-supervised learning or reinforcement learning, could improve the model’s adaptability and robustness in various scenarios. While this study focuses on SJIs in swimmers, future research could extend to other types of pathological conditions, such as knee osteoarthritis or ligament injuries. By applying the MSFFN model to various medical imaging tasks, its broad applicability and clinical value can be further validated. Author contributions Lina Dai: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparationMd Gapar Md Johar: writing—review and editing, visualization, supervision, project administration, funding acquisitionMohammed Hazim Alkawaz: methodology, software, validation, formal analysis, investigation. Funding The authors are grateful to the Guangzhou College of Commerce in China, Management and Science University in Malaysia, and the University of Mosul in Iraq for their supports. This work is funded by Guangdong Provincial Education Science Foundation [grant number: 2023GXJK407], the Scientific Research Foundation of Guangzhou [grant number: 202201011667], the Philosophy and Social Science Foundation of Guangzhou [grant number: 2021GZGJ145], the Guangdong Province Undergraduate Teaching Quality and Teaching Reform Engineering Construction Project (2022SJJXGG992), the school level scientific research project of Guangzhou Business School (2023XJYB33). Data availability The datasets used and/or analyzed during the current study are available from the corresponding author Lina Dai on reasonable request via e-mail dailina27@163.com. Although the training model has not yet been publicly released, we are considering the following measures to enhance the reproducibility of our research and contribute to the community in future studies: Model sharing: we plan to publish the code and weights of the trained model in a public GitHub repository when appropriate, allowing other researchers to access and utilize them.Detailed documentation: if we choose not to release the model at this time, we will ensure that the GitHub repository includes clear instructions for other researchers on how to obtain the model weights or build the model, including necessary environment setup and code examples.Ongoing updates: we commit to continuously improving and updating the availability of data and the model based on feedback, ensuring it provides practical value to the broader scientific community. Competing interests The authors declare no competing interests. Ethics statement The studies involving human participants were reviewed and approved by School of Information Technology and Engineering, Guangzhou College of Commerce Ethics Committee (Approval Number: 2022.29384384). The participants provided their written informed consent to participate in this study. All methods were performed in accordance with relevant guidelines and regulations. 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 The datasets used and/or analyzed during the current study are available from the corresponding author Lina Dai on reasonable request via e-mail dailina27@163.com. Although the training model has not yet been publicly released, we are considering the following measures to enhance the reproducibility of our research and contribute to the community in future studies: Model sharing: we plan to publish the code and weights of the trained model in a public GitHub repository when appropriate, allowing other researchers to access and utilize them.Detailed documentation: if we choose not to release the model at this time, we will ensure that the GitHub repository includes clear instructions for other researchers on how to obtain the model weights or build the model, including necessary environment setup and code examples.Ongoing updates: we commit to continuously improving and updating the availability of data and the model based on feedback, ensuring it provides practical value to the broader scientific community.