ENHANCED MULTI-SCALE ATTENTION-BASED 3D HYBRID DEEP NETWORK (EMA-3DNET) – AN IMPROVED METHODOLOGY FOR SPORTS KNEE INJURY DETECTION USING MRI

Abstract

Accurate and early detection of knee injuries is critical in sports medicine to ensure timely treatment and prevent long-term disability. Magnetic Resonance Imaging (MRI) is the gold regular for non-invasive analysis of musculoskeletal injuries, especially anterior cruciate ligament (ACL) and meniscus tears. Traditional deep learning approaches often fail to exploit the full spatial and contextual relationships embedded in volumetric MRI data, limiting their clinical utility. In this study, we introduce anEnhanced Multi-Scale Attention-Based 3D Hybrid Deep Network (EMA-3DNet)designed to overcome these limitations through multi-scale feature extraction, 3D convolutional encoding, and channel-spatial attention mechanisms.EMA-3DNet integrates a 3D ResNet-based backbone with a Feature Pyramid Network (FPN) and a Convolutional Block Attention Module (CBAM) or Transformer-based concentration, enabling simultaneous classification and segmentation of injuries. Extensive evaluations were conducted using the MRNet and Osteoarthritis Initiative (OAI) datasets. The proposed model achieved a classification accuracy of 96.1%, Dice similarity coefficient of 0.91, and a significant performance improvement over existing 2D CNN and plain 3D CNN architectures.The model offers high clinical interpretability by producing overlay segmentation maps highlighting injury-prone regions. This not only enhances diagnostic precision but also provides critical insights to orthopedic specialists and sports physicians. EMA-3DNet presents a novel direction in integrating 3D attention-driven networks with medical imaging workflows and paves the way for future developments in AI-assisted radiological diagnostics.