Explainable Multi-planar Multi-slice Learning for Efficient 3D MRI Knee Osteoarthritis Classification

Accurate multi-class grading of knee osteoarthritis (OA) from magnetic resonance imaging (MRI) remains challenging due to subtle inter-class differences and high structural similarity between adjacent Kellgren-Lawrence (KL) grades. While deep convolutional neural networks have shown promise in OA classification, existing approaches often rely on large parameter counts and primarily focus on binary discrimination. In this work, we propose a lightweight hybrid framework for five-class KL grading from volumetric knee MRI. The proposed architecture integrates 3D convolutional feature extraction, complementary multi-plane 2D representations, and a transformer-based module for global contextual modeling across slices. Comprehensive experiments on the Osteoarthritis Initiative dataset were conducted against established 3D ResNet and DenseNet variants as well as prior multi-class OA classification methods under identical training settings. The proposed model achieves the highest overall accuracy of 66.08% and the best weighted AUC of 0.8740 while using only 1.1M parameters, substantially fewer than competing architectures. Ablation studies confirm the complementary contributions of volumetric modeling, multi-plane feature extraction, and transformer-based global reasoning. Furthermore, SHAP-based visualization demonstrates that the model focuses on clinically relevant anatomical regions, including cartilage surfaces and joint space areas. These results indicate that the proposed framework provides an effective and parameter-efficient solution for multi-class KL grading in knee MRI.