EMA-3DNet++: Federated Attention-Based 3D Hybrid Network for Intelligent Football Knee MRI Analysis

Abstract

Football players are highly susceptible to knees harms, particularly ACL damages, meniscus damage, in addition to cartilage degeneration due to rapid directional changes, high-impact tackles, and repetitive stress. Untimely recognition moreover severity grading are critical for reducing enduring performance loss and preventing career-threatening conditions. This study proposes an improved Multi-Scale Attention-Based 3D Hybrid Deep Network (EMA-3DNet++) tailored for football damage assessment utilizing volumetric knee MRI data. The model integrates multi-scale 3D convolutional feature extraction, attention-driven contextual refinement, transformer-based global representation learning, and graph-based anatomical structure modeling. The proposed system performs simultaneous segmentation of knee structures, injury classification, and severity grading within a unified multi-task learning framework. Additionally, federated learning enables collaborative training across sports hospitals and football medical centers without sharing sensitive athlete data. Explainable attention maps provide interpretable insights for sports physicians and team medical staff. Experimental evaluation demonstrates superior segmentation accuracy, robustness to imaging variations, and reduced inference time compared to conventional 3D deep learning architectures. The system supports real-time injury assessment during tournament seasons and pre-season screening programs. The proposed framework offers a clinically deployable AI solution for professional football injury management, enabling precision diagnosis, faster rehabilitation planning, and improved long-term athlete safety.