Deep Learning-Based Knee Osteoarthritis MRI Image Classification Method

In the classification of knee osteoarthritis (KOA) using MRI images, the application of end-to-end deep learning techniques holds significant importance. The raw knee joint images contain multiple complex attributes, including shape, size, and spatial position, all of which critically influence the performance of neural networks in KOA classification. However, knee joint imaging often exhibits intricate anatomical structures and subtle variations, particularly when different categories share high similarities. Traditional methods face limitations in extracting fine-grained features, making it challenging to distinguish key discriminative patterns among different classes and increasing classification difficulty. Moreover, conventional convolutional neural networks (CNNs) primarily rely on a hierarchical stacking of layers, where single convolutional kernels focus on capturing local features at the input stage. This approach frequently results in the loss of critical fine details within images, leading to suboptimal classification performance. To address the challenges of inadequate local feature extraction and class imbalance in KOA MRI classification, this study introduces a classification network based on the Swin Transformer. By integrating multiple novel modules and optimization strategies, the proposed model aims to enhance classification accuracy and improve generalization performance.