Deep learning-based automated segmentation and quantification of glenoid and humeral head defects

Abstract

PURPOSE: Humeral head and glenoid bone defects resulting from recurrent shoulder dislocation significantly compromise joint stability and complicate surgical planning. Conventional manual assessment is subjective and labor-intensive. This study aims to develop and validate an automated, deep learning-based deep defect detection network for the simultaneous segmentation and precise quantification of these defects using edge-global MRI. METHODS: A retrospective study was conducted using 800 glenoid and 792 humeral head MRI scans from a single clinical center between June 2024 and August 2025. The proposed deep defect detection network architecture utilizes you only look once version 8-segment as a shared backbone, integrating specialized modules: an edge-global attention module for boundary refinement, an ASFFHead for multi-scale feature fusion, an optimized vision transformer block for efficient feature extraction, and a dynamic mask generation module to handle complex defect morphologies. The dataset was partitioned into training, validation, and test sets (8:1:1 ratio). Performance was evaluated using intersection over union, dice similarity coefficient, precision, and Hausdorff distance. Statistical comparisons were performed against established models, including U-Net, region-based convolutional neural network region-based convolutional neural network, U-NeXt, you only look once version 8, and you only look once version 11. RESULTS: The proposed model achieved superior segmentation performance, with a mean IoU of 89.77% and a dice score of 94.71%. In boundary delineation, the model recorded a Hausdorff distance of 2.3 mm, significantly outperforming the baseline you only look once version 8 (2.6 mm) and U-Net (3.1 mm). Ablation studies confirmed that the dynamic mask generation module and edge-global attention module provided the highest gains in accuracy, particularly for small-scale and irregular defect regions. The defect area calculation algorithm demonstrated high consistency with expert annotations, facilitating the classification of defect severity (mild, moderate, severe) based on clinical thresholds. CONCLUSION: The deep learning-based joint diagnostic network provides an efficient and precise tool for automated shoulder defect quantification. By integrating edge-sensitive attention and dynamic weight adjustment, the model effectively addresses the challenges of low contrast and morphological complexity in MRI. This approach offers robust support for personalized surgical planning and objective clinical assessment in orthopedic diagnostics.

Ähnliche Arbeiten