The protocol of this retrospective study was approved by the Institutional Review Board of Taipei Veterans General Hospital (TVGH; IRB number: 2023-01-024AC). The requirement for informed consent was waived owing to the retrospective nature of the study and the use of deidentified imaging data. All procedures were conducted in accordance with the ethical standards of the institutional research committee and the principles outlined in the Declaration of Helsinki. The study also complied with relevant national legislation and institutional guidelines governing human subject research. Participants did not receive any compensation for this study.

Coronal T2-weighted shoulder MRI images from 1192 cases were collected for training the new AI-based hierarchical model developed in this study. These images were predominantly obtained from TVGH, but the dataset also included images collected across 65 institutions or uploaded from external facilities for second-opinion evaluations.

Patients with tumors, prior shoulder surgery, humeral head osteonecrosis, moderate to severe osteoarthritis (Kellgren-Lawrence [KL] grades 2 to 4), or a history of fractures around the shoulder were excluded. KL grading was assessed using anteroposterior shoulder radiographs, based on osteophyte formation and joint space narrowing in accordance with established definitions. The distinction between KL grades 1 and 2 followed the original criteria, with definite osteophyte formation serving as the primary threshold for grade 2 [14]. Based on this definition, patients with KL grade ≥2 were excluded to minimize potential confounding from glenohumeral degenerative changes. In this study, KL grading was applied as a pragmatic radiographic exclusion criterion rather than for formal characterization of shoulder osteoarthritis. Although originally developed for other joints, KL grading has been applied in prior radiographic and population-based studies of glenohumeral osteoarthritis, with demonstrated interobserver reliability [15,16].

Cases were categorized into the following 3 groups on the basis of supraspinatus tendon status: intact tendons (n=449), tendinopathy or partial-thickness tears (n=235), and full-thickness tears (n=508). The reference standard for labeling intact tendons and tendinopathy/partial-thickness tears was based on MRI interpretations by consensus between 1 musculoskeletal radiologist and 2 orthopedic surgeons, each with over 10 years of specialty training experience. Full-thickness tears were confirmed through consistent MRI findings and verification by arthroscopic surgical reports.

Of note, tendinopathy and partial-thickness tears were grouped into a single category based on both clinical and methodological considerations. Structurally, tendon continuity is preserved in both these conditions, in contrast to full-thickness tears, which disrupt tendon integrity. Clinically, the management of tendinopathy and partial-thickness tears is similar, with conservative treatment (rest, rehabilitation, medication, or injections) typically preferred before considering surgical intervention. Moreover, the imaging findings of tendinopathy and early partial-thickness tears often overlap, and interobserver agreement in distinguishing them can be limited [17]. Combining these 2 entities therefore reduces diagnostic variability and aligns more closely with clinical practice.

Figure 1 presents the different supraspinatus tendon pathologies, including intact tendons, tendinopathy/partial-thickness tears, and full-thickness tears.

MRI-based diagnostic criteria are as follows: tendinopathy is defined by tendon thickening with an inhomogeneous signal intensity and increased signal across all pulse sequences; partial-thickness tears are defined by fluid signal intensity partially filling a gap within the tendon, most evident on fat-suppressed T2-weighted images; and full-thickness tears are defined by complete disruption of the tendon, appearing as high signal intensity spanning the entire tendon thickness on all pulse sequences [18,19].

Cases from each year between 2018 and 2024 were stratified by acquisition year and randomly assigned at the patient level to a training set (64%), a validation set (16%), and an internal test set (20%), ensuring that the year-wise distribution was preserved across all subsets.

The training set consisted of 287 intact tendon cases, 150 tendinopathy or partial-thickness tear cases, and 327 full-thickness tear cases. The validation set included 72 intact tendon cases, 38 tendinopathy or partial-thickness tear cases, and 79 full-thickness tear cases. The internal test set comprised 90 intact tendon cases, 47 tendinopathy or partial-thickness tear cases, and 102 full-thickness tear cases.

The training, validation, and internal test sets demonstrated similar mean echo time and repetition time values. Age and sex distributions were similar across the training, validation, and internal test sets. The distributions of shoulder side and tear type were also comparable among the datasets (Table 1).

A total of 1192 cases were collected, and the internal cohort was predominantly obtained from TVGH (987 cases; Table 2).

For the internal cohort, the dataset included MRI scans acquired using 29 scanner models from the following 6 vendors: Canon Medical Systems (Ōtawara-shi, Tochigi, Japan), GE Healthcare (Chicago, IL), Philips Medical Systems (Andover, MA; headquartered in Amsterdam, Netherlands), Siemens Healthineers (Forchheim, Germany), Hitachi Medical Corporation (Tokyo, Japan), and Toshiba Medical Systems (Ōtawara-shi, Tochigi, Japan) (Table 1).

The in-plane resolution (X-Y pixel spacing) ranged from approximately 0.19 mm to 0.86 mm, with the majority of cases falling within the range of 0.30‐0.39 mm. Within this range, 0.3125 mm was the most frequently observed in-plane resolution (496 cases). Slice spacing ranged from 3.0 mm to 5.0 mm, with 3.5 mm being the most commonly used spacing. Slice thickness varied from 2.0 mm to 4.5 mm, and the most common thickness was 3.0 mm. The detailed distributions of slice thickness, slice spacing, and in-plane resolution across the training, validation, and internal test sets are summarized in Tables S1 and S2 in Multimedia Appendix 1.

An independent external validation dataset was created from institutions not included in the model development cohort to evaluate the generalizability of the proposed model.

A total of 89 cases were collected from 12 medical institutions across Taiwan, with the largest contribution from Taoyuan Hospital, Ministry of Health and Welfare (37 cases; Table 3). The external validation set comprised 35 intact tendon cases, 11 tendinopathy or partial-thickness tear cases, and 43 full-thickness tear cases.

None of the external cases were used during model training, validation, or hyperparameter tuning (Table 1).

A 3D ResNet-18 CNN was constructed for classification. The model architecture is a spatiotemporal extension of the conventional ResNet framework, incorporating residual connections and 3D convolutions to process volumetric data effectively [20]. The 3D ResNet-18 model used in this study was based on the torchvision library implementation within the PyTorch framework (Figure 2). The model used standard BatchNorm3d layers as implemented in the default architecture.

Using transfer learning, the model was initialized with pretrained weights, and the final fully connected layer was modified to perform 2-class classification for each task. During training, a validation set was used to monitor performance and select the best-performing model. Score-CAM was used to generate heatmaps for model interpretability, highlighting the most influential regions in model predictions.

To achieve 3-class classification, a hierarchical model composed of 3 sequential 3D ResNet-18 classifiers was designed as follows (Figure 3):

The use of this stepwise approach helped reduce classification complexity and improve overall performance by leveraging domain knowledge of clinical relevance.

Although the final prediction was reported as a 3-class outcome (intact tendons, tendinopathy/partial-thickness tears, and full-thickness tears), the proposed framework performed hierarchical inference through 2 sequential binary classifiers. Specifically, the first-stage model distinguished full-thickness tears from non–full-thickness cases, and only cases predicted as non–full-thickness were subsequently passed to the second-stage model to differentiate intact tendons from tendinopathy/partial-thickness tears. Therefore, the reported 3-class confusion matrices reflected the final outputs of a hierarchical decision process rather than a flat 3-class classifier.

In addition, for probabilistic evaluation (receiver operating characteristic [ROC] curves and macro area under the curve [AUC]), we derived a unified 3-class probability distribution using conditional probability decomposition. Specifically, the predicted conditional probabilities were defined as follows:

where ITP denotes tendinopathy/partial-thickness tears.

Final 3-class probabilities were computed using conditional probability decomposition as follows:

P (intact∣𝑥) = P (non–full-thickness∣𝑥)⋅ P (intact∣non–full-thickness, 𝑥)

P (ITP∣𝑥) = P (non–full-thickness∣𝑥)⋅ P (ITP∣non–full-thickness, 𝑥)

P (full-thickness∣𝑥) = P (full-thickness∣𝑥)

where probabilities from model F and model ITP were obtained via Softmax normalization of the corresponding logits. This formulation ensures that the final outputs form a valid probability distribution that sums to 1.

These derived probabilities were used to compute one-vs-rest ROC curves and macro AUC for the 3-class classification task.

The dataset exhibited class imbalance across diagnostic categories, particularly for partial-thickness tears and tendinopathy. In the proposed hierarchical framework, class weighting, a strategy shown by Buda et al [21] to be effective and stable for handling imbalanced medical imaging datasets in CNNs, was applied only to the second-stage classifier (model ITP), which distinguishes intact tendons from partial-thickness tears and tendinopathy.

This design choice was motivated by the observation that class imbalance and visual ambiguity are most pronounced at this stage. In contrast, the first-stage classifier (model F), which detects full-thickness tears, targets a clinically distinct and more visually separable category. Selective application of class weighting therefore allowed us to improve sensitivity for the most challenging minority class without destabilizing optimization in the first-stage model as follows:

To visualize the regions of interest that contributed most to model decisions, Score-CAM was implemented [22]. Unlike gradient-based methods, such as Gradient-weighted Class Activation Mapping [23], Score-CAM assesses the contribution of each activation map by forwarding masked images and observing their effect on the model’s output rather than relying on backpropagated gradients. This method provides intuitive, spatially localized visualizations of model attention to support interpretability. The evaluation of these visualizations was qualitative and based on representative cases, without formal quantitative assessment or radiologist-based validation (Figures S1-3 in Multimedia Appendix 2).

For comparison, we trained a conventional flat 3-class classifier to predict intact tendons, tendinopathy/partial-thickness tears, and full-thickness tears directly. This model used the same preprocessing pipeline as the hierarchical framework, including left-right shoulder classification for anatomical standardization, followed by selection of the central 16 slices, resizing, and intensity normalization. The same 3D ResNet-18 backbone, training-validation-test splits, and optimization settings were applied.

To mitigate class imbalance, class-weighted cross-entropy loss was used during training, with class weights computed from the training set using the balanced weighting scheme provided by the scikit-learn library.

The only architectural difference was that the flat baseline used a single 3-class Softmax output layer, replacing the 2-stage hierarchical design.

All models were trained for 1000 epochs on a single GPU. Because wall-clock training time was not explicitly recorded during the full training process, GPU training time was measured over 10 epochs using identical hardware, batch size, and training settings. The total training time for 1000 epochs was then extrapolated assuming linear scaling with respect to the number of epochs.

For multiclass classification, precision, F₁-score, sensitivity, and specificity were computed using macro-averaging, whereby each class (intact tendons, tendinopathy/partial-thickness tears, and full-thickness tears) contributes equally regardless of class prevalence.

ROC AUC values were computed using a one-vs-rest strategy from Softmax probability outputs and have been reported as macro-averaged AUC values. Overall accuracy corresponded to micro-averaged performance.

The impact of different preprocessing strategies on model performance was evaluated using model ITP under a controlled ablation framework. Balanced accuracy (macro-averaged sensitivity) was used as the primary evaluation metric to account for class imbalance. The network architecture, dataset split, and training protocol were identical across all experiments, with only preprocessing components being varied.

As summarized in Table 4, applying interpolation-based slice standardization and voxel spacing normalization without left-right mirroring resulted in a balanced accuracy of 0.815. Introducing left-right mirroring while using zero-padding to standardize slice count yielded a modest increase in balanced accuracy to 0.820. This improvement suggests that anatomical orientation alignment reduces directional variability; however, the use of padding may introduce nonphysiological discontinuities along the slice dimension.

When left-right mirroring and interpolation were applied without voxel spacing normalization, balanced accuracy decreased to 0.799. This reduction indicates that heterogeneity in acquisition parameters, including in-plane resolution and slice spacing, negatively affects model robustness when spatial normalization is omitted.

The highest balanced accuracy (0.826) was achieved when left-right mirroring, interpolation-based slice standardization, and voxel spacing normalization were jointly applied. Based on these findings, the full preprocessing pipeline was selected for subsequent experiments and applied consistently to both model ITP and model F.

All models were trained for 1000 epochs on a single GPU. As wall-clock time was not recorded for the full training run, runtime was measured over 10 epochs using end-to-end wall-clock timing (including both training and validation phases, as well as data loading and evaluation overhead) and then extrapolated to 1000 epochs assuming linear scaling.

Based on this estimation, the total end-to-end GPU runtime for 1000 epochs was approximately 41.9 hours for the left-right classifier, 72.5 hours for model F, and 22.4 hours for model ITP (Table 5). These values should be interpreted as estimated runtimes, as linear extrapolation may not fully capture potential variability in longer training runs (Table 5).

Shoulder laterality (left or right) was determined after first training a binary classifier. The model achieved perfect validation performance (100%) by epoch 2.

Model F was then trained to differentiate full-thickness tears from non–full-thickness cases. The highest validation-balanced accuracy of 96.74% was achieved at epoch 93. Finally, model ITP was trained to distinguish intact tendons from cases with tendinopathy or partial-thickness tears. This model reached its best validation-balanced accuracy of 87.10% at epoch 983.

Training progress and performance metrics across epochs for each model are presented in Figure S4 in Multimedia Appendix 2.

Model F, designed to differentiate full-thickness tears from non–full-thickness cases, achieved an accuracy of 93.7% on the internal validation set. The model demonstrated a macro precision of 93.5% and a macro F₁-score of 93.6%. The sensitivity for full-thickness tears was 94.1%, and the specificity for non–full-thickness cases was 93.4%. The ROC AUC was 0.984.

Model ITP, which further classified non–full-thickness cases into intact tendons or tendinopathy/partial-thickness tears, achieved an accuracy of 82.6%. The model yielded a macro precision of 80.8% and a macro F₁-score of 81.3%. The sensitivity for tendinopathy/partial-thickness tears was 83.0%, while the specificity for intact tendons was 82.3%. The ROC AUC was 0.904.

The hierarchical model, which used the left-right model for preprocessing and integrated outputs from model F and model ITP for the final 3-class classification, achieved an accuracy of 84.1% and a macro-average AUC of 0.942. The macro sensitivity and macro specificity were 81.0% and 92.4%, respectively. In addition, the model yielded a macro precision of 80.7% and a macro F₁-score of 80.6% (Figure 4A; Tables 6-8).

On the external validation set, the left-right model maintained perfect performance, achieving 100% accuracy, precision, sensitivity, and specificity, and an ROC AUC of 1.000, confirming strong generalizability for shoulder laterality identification.

Model F achieved an accuracy of 87.7% on the external validation set. The model demonstrated a macro precision of 88.3% and a macro F₁-score of 87.7%. The sensitivity for full-thickness tears was 93.0%, while the specificity for non–full-thickness cases was 82.8%. The ROC AUC was 0.95, indicating good discrimination despite reduced performance compared with internal validation.

Model ITP showed increased performance variability on the external dataset, achieving an accuracy of 76.2%. The macro precision and macro F₁-score were 69.4% and 69.8%, respectively. The sensitivity for tendinopathy/partial-thickness tears was 64.0%, while the specificity for intact tendons was 80.0%. The ROC AUC was 0.815.

The hierarchical model, which used the left-right model for preprocessing and integrated outputs from model F and model ITP for the final 3-class classification, achieved an accuracy of 76.4% and a macro-average AUC of 0.893. The corresponding macro sensitivity, macro specificity, macro precision, and macro F₁-score were 68.1%, 87.7%, 68.3%, and 67.4%, respectively (Figure 4B; Tables 6-8).

This study describes a hierarchical deep learning framework for classifying supraspinatus tendon pathologies on coronal T2-weighted shoulder MRI. By separating the diagnostic task into 2 clinically guided stages—first detecting full-thickness tears and then distinguishing intact tendons from partial-thickness tears or tendinopathy—the model aligns with real-world diagnostic workflow patterns. The hierarchical design achieved high sensitivity for full-thickness tears and showed consistent performance on an external dataset, supporting its generalizability. Compared with a conventional flat model, the hierarchical approach was associated with numerically improved performance in selected metrics; however, the 95% bootstrapped CIs overlapped, suggesting that these differences should be interpreted with caution. Additionally, integration of Score-CAM enhanced model interpretability by highlighting clinically relevant features that informed model predictions. These results suggest that the AI-based framework described herein may support radiologists and orthopedic surgeons in improving diagnostic efficiency, consistency, and confidence in musculoskeletal imaging.

Recent advances in deep learning have correspondingly advanced the automated analysis of shoulder MRI for diagnosing rotator cuff tears. Several notable studies have explored various model architectures and approaches. Lin et al [24] developed an ensemble model using 4 parallel 3D ResNet-50 networks to classify rotator cuff tears as no tear, partial-thickness tear, or full-thickness tear. Guo et al [25] introduced a 2D CNN model based on the Xception architecture for the automated diagnosis of supraspinatus tears. Lee et al [26] used a YOLO v8–based model to detect rotator cuff tears across axial, coronal, and sagittal planes. Esfandiari et al [27] proposed a custom 2D CNN for binary classification of rotator cuff tears versus healthy cases. Yao et al [28] developed a 3-stage deep learning pipeline incorporating slice selection (ResNet), segmentation (U-Net), and a multi-input CNN classifier for supraspinatus tear detection.

While these prior studies demonstrated promising results, most were limited by single-center datasets and variable sample sizes—ranging from 100 to 11,000 MRI examinations—which may constrain generalizability and clinical applicability. In addition, prior studies differed substantially in class definitions, input configurations (eg, single-plane vs multisequence MRI), and imaging vendor characteristics, further limiting direct comparison across studies and hindering the assessment of real-world performance.

Recent work has demonstrated the feasibility of code-free deep learning (CFDL) platforms for musculoskeletal MRI analysis, including the detection and classification of supraspinatus tendon pathologies on shoulder MRI, enabling rapid model development without manual coding [29]. Similar advantages of CFDL approaches, such as accessibility and ease of use, have also been reported in other radiologic applications, including chest radiograph classification, detection, and segmentation [30].

However, code-free frameworks generally provide limited flexibility in task-specific preprocessing, explicit volumetric modeling, and structured decision strategies, which are critical for hypothesis-driven evaluation and clinical interpretability. In this study, a coded deep learning framework was therefore adopted to enable explicit 3D preprocessing, hierarchical task decomposition, and transparent analysis of sensitivity-specificity tradeoffs and error propagation. This design aligns with the clinical diagnostic workflow of rotator cuff evaluation and supports interpretable, clinically grounded model assessment. We did not adopt 2D slice-level aggregation approaches, as rotator cuff tear evaluation inherently requires the assessment of 3D spatial continuity across MRI slices. Clinically relevant features, such as tear extent, morphology, and tendon retraction, are distributed across multiple slices and cannot be reliably captured using isolated 2D representations, which may lead to suboptimal diagnostic performance.

Building on earlier work, this study has presented a hierarchical deep learning framework trained on 1192 shoulder MRI scans from a dataset that was predominantly derived from a single institution but also contained images collected across 65 institutions and acquired using 29 different scanner models. This dataset reflects some degree of real-world clinical variability and provides preliminary evidence regarding the robustness and generalizability of the model. Additionally, the new model integrates Score-CAM to provide intuitive, heatmap-based visualizations of the regions contributing to each prediction, supporting interpretability and offering clinical utility by assisting radiologists in verifying the model’s diagnostic rationale.

Although the classification task nominally involves only 3 diagnostic categories, the challenge lies not in the number of classes but in the substantial visual overlap between adjacent categories, particularly among intact tendons, partial-thickness tears, and early tendinopathy. In our flat 3-class baseline, this intrinsic ambiguity resulted in consistently low sensitivity for tendinopathy/partial-thickness tears, especially on the external cohort, despite using the same backbone architecture, preprocessing pipeline, and training configuration as in the hierarchical framework.

To address this limitation, we adopted a hierarchical classification strategy composed of 2 sequential binary decision steps, motivated by task decomposition rather than increased architectural complexity. In this framework, model F first identifies full-thickness tears and excludes them from further consideration, and then, model ITP classifies the remaining cases as intact or tendinopathy/partial-thickness tears. By explicitly separating a clinically high-priority decision from the assessment of more subtle pathology, the hierarchical design simplifies individual decision boundaries and closely mirrors diagnostic reasoning in routine clinical practice.

Quantitatively, the hierarchical model demonstrated numerically higher balanced performance compared with the flat 3-class baseline across both internal and external cohorts. Balanced accuracy increased from 77.1% to 81.0% on the internal test set and from 64.5% to 68.1% on the independent external cohort. However, the 95% bootstrapped CIs overlapped; therefore, these numerical differences should be interpreted cautiously and may not represent statistically significant differences. These differences were primarily driven by higher sensitivity for tendinopathy/partial-thickness tears—the most diagnostically challenging category—particularly in the external cohort (hierarchical vs flat: sensitivity 45.5% vs 18.2%). This represents a trend toward improved detection in this class, although the difference may not be statistically significant. This increase in sensitivity was accompanied by more false-positive predictions among intact cases, reflecting a sensitivity-specificity tradeoff. This tradeoff was explicitly characterized through per-class metrics and confusion matrices.

As expected, hierarchical models are susceptible to error propagation, whereby misclassification at an earlier decision stage cannot be corrected downstream. Our leakage analysis suggests that a substantial proportion of residual errors arose from clinically recognized ambiguity between intact tendons and tendinopathy/partial-thickness tears, rather than widespread routing failures in model F. Compared with the flat classifier, the hierarchical framework was associated with higher sensitivity for subtle tendon pathology, accompanied by reduced specificity for intact cases. This reflects a sensitivity-specificity tradeoff that may be clinically relevant; however, given the overlapping 95% bootstrapped CIs, these differences did not reach statistical significance. These findings should therefore be interpreted as a trend rather than a definitive improvement.

Finally, the 2-stage hierarchical architecture, together with the preliminary left-right shoulder classification used to ensure anatomical consistency of inputs, may enhance both interpretability and clinical plausibility. In addition to architectural design, several preprocessing strategies informed by imaging and clinical domain knowledge were incorporated and may have contributed to model performance.

The hierarchical framework was associated with numerically improved selected classification metrics and enhanced interpretability—an important factor for clinical integration. However, given the overlapping 95% bootstrapped CIs, these performance differences did not reach statistical significance. These findings suggest that structuring AI models in alignment with established diagnostic workflows may be beneficial, particularly in scenarios where certain pathology categories (eg, full-thickness tears) are visually distinct and diagnostically prioritized. Nevertheless, these observations should be interpreted as indicative of a trend, and further validation with larger cohorts and formal statistical testing is warranted.

The integration of Score-CAM further contributed to model transparency by generating interpretable heatmaps. Notably, attention maps from the left-right classifier and model F closely corresponded to anatomical regions of interest typically emphasized by experienced clinicians, suggesting that these models learned clinically relevant features. However, model ITP occasionally attended to regions inconsistent with the surgical focus, indicating a need for further refinement. Improving the model’s sensitivity to subtle findings, such as tendinopathy and partial-thickness tears, will likely require a larger, well-annotated dataset with richer representation of these pathologies. Notably, the evaluation of Score-CAM in this study was qualitative and based on representative examples rather than a systematic or quantitative assessment; therefore, these visualizations should be interpreted as supporting interpretability rather than providing definitive evidence of accurate anatomical localization.

This study has several strengths. First, although the internal cohort was predominantly derived from a single institution, it also included MRI data from multiple external sources, providing an opportunity to preliminarily assess model performance across heterogeneous imaging settings. Second, the hierarchical classification strategy mirrored real-world diagnostic reasoning, contributing to numerically improved selected performance metrics and clinical interpretability.

This study also has several limitations. First, although volumetric slice standardization and manual quality control were implemented to ensure supraspinatus tendon coverage, it remains possible that peripheral or atypically positioned tear extensions were not fully captured within the standardized 16-slice volume. This design choice balanced anatomical coverage and computational feasibility. Future work may explore adaptive slice selection or attention-based volumetric models. Second, the external validation cohort—while useful for assessing generalizability—contained only 11 cases of tendinopathy or partial-thickness tears. This limited subgroup sample size introduced statistical fragility and wide CIs, restricting the strength of conclusions regarding external generalizability for this category. Sensitivity estimates for tendinopathy/partial-thickness tears should therefore be interpreted with caution, and larger external cohorts will be required to verify performance for this subgroup. Third, potential label bias may exist, as full-thickness tears were confirmed using surgical or arthroscopic findings, while partial-thickness tears and tendinopathy were labeled based on imaging consensus. This difference in reference standards may introduce classification bias. Accordingly, the model’s performance for partial-thickness tears and tendinopathy reflects agreement with expert interpretation rather than true diagnostic accuracy against a surgical gold standard, which may limit the achievable performance for these classes. Fourth, the hierarchical framework may be susceptible to cumulative error propagation, in which misclassifications at early stages influence downstream outputs. This tradeoff reflects an intentional prioritization of sensitivity for partial-thickness tears over global accuracy but may reduce specificity for certain subgroups. Fifth, the model was trained exclusively on coronal T2-weighted sequences, whereas ground truth labels were established based on full multiplanar and multisequence clinical MRI review. This modality gap may limit the recognition of subtle partial-thickness tears, which often benefit from sagittal or proton-density fat-suppressed views. Future work incorporating multisequence or multiplane inputs may enhance diagnostic performance. Sixth, tendinopathy and partial-thickness tears were grouped into a single category to reflect overlapping early management in routine clinical practice. However, this grouping reduces granularity for surgical planning, particularly around thresholds for operative referral versus conservative management. Seventh, patients with moderate-to-severe glenohumeral osteoarthritis (KL grades 2‐4) were excluded, which may limit applicability to older individuals at the highest risk for degenerative rotator cuff pathology and introduce a form of selection bias. Prospective validation in more heterogeneous clinical populations is warranted. Eighth, with respect to the left-right classification task, although all MRI volumes were visually inspected to confirm the absence of burned-in laterality markers or textual overlays, the possibility that subtle nonanatomical cues contributed to the classification cannot be completely excluded. Ninth, the assessment of Score-CAM was qualitative and lacked formal radiologist-based or quantitative validation, and thus, the extent to which the highlighted regions correspond to true anatomical focus remains uncertain. Tenth, most cases originated from a single hospital, TVGH, which may introduce center-specific imaging bias and limit the model’s generalizability to other institutions. Finally, alternative model architectures, including 2D slice-level ensembles or transformer-based volumetric models, were not explored in order to isolate the effect of hierarchical diagnostic task structuring. Future studies may investigate hybrid approaches to further align model behavior with clinical decision-making pathways.

In this study, we developed a hierarchical deep learning framework to classify supraspinatus tendon pathologies on shoulder MRI. The framework decomposed the diagnostic task into two binary stages—first distinguishing full-thickness tears and then differentiating intact tendons from tendinopathy or partial-thickness tears—reflecting clinical diagnostic reasoning. Across internal and external evaluations, the hierarchical approach showed numerically improved recognition of tendinopathy and partial-thickness tears relative to a conventional flat 3-class model while maintaining high sensitivity for full-thickness tears. These findings suggest that clinically motivated task structuring may support more reliable supraspinatus pathology classification and may align with radiological interpretation workflows, although further validation with larger cohorts and formal statistical testing is warranted.