Enhanced Multi-Scale Attention 3D Deep Network (EMA-3DNet++) for Segmentation, Classification, and Severity Grading of Knee Injuries

Abstract

Sports-related knee injuries such as anterior cruciate ligament (ACL) tears, meniscal degeneration, and cartilage defects demand precise and early diagnosis to prevent long-term functional impairment. Although deep learning–based medical image analysis has significantly improved automated knee assessment, existing 3D convolutional models often suffer from limited global contextual reasoning, inadequate cross-structure modeling, and poor generalization across heterogeneous clinical datasets. To address these limitations, this research work proposes EMA-3DNet++, a next-generation Federated Multi-Scale Transformer–Graph Hybrid 3D Deep Network designed for accurate, explainable, and privacy-preserving sports knee analysis. EMA-3DNet++ integrates self-supervised pre-trained 3D encoders with multi-scale convolutional feature extraction and Swin Transformer blocks to capture both fine-grained anatomical details and long-range spatial dependencies in volumetric MRI data. A Graph Neural Network (GNN) module explicitly models structural relationships among knee components—ACL, PCL, meniscus, cartilage, tibia, and femur—enhancing multi-label injury classification and structural consistency. To further improve robustness, a hybrid temporal refinement unit incorporating LSTM and energy-efficient spiking neural layers captures motion dynamics in sequential scans. The framework adopts a multi-task learning strategy, simultaneously performing segmentation, injury classification, and severity grading.

To overcome data scarcity and privacy constraints, EMA-3DNet++ supports federated learning with secure aggregation, enabling collaborative training across institutions without data sharing. An advanced explainability layer combining 3D Grad-CAM++, attention rollout, and uncertainty estimation enhances clinical interpretability. Experimental evaluation on benchmark datasets including MRNet, SKI10, and OAI demonstrates that EMA-3DNet++ achieves superior Dice coefficient (94–96%), improved sensitivity and specificity, and enhanced cross-domain generalization compared to state-of-the-art 3D CNN baselines. The proposed framework represents a scalable, clinically deployable, and energy-efficient solution for next-generation AI-assisted sports knee diagnostics.