A deep learning-based approach to knee joint semantic segmentation and diagnosis prediction on 3D MRI images

Accurate and timely diagnosis of meniscus tears is crucial for patient recovery. This study proposes a fully automated deep learning-based approach for knee joint semantic segmentation and meniscus tear diagnosis, designed to provide clinicians with a visual model of the knee joint and improve diagnostic accuracy and efficiency for meniscus injuries. First, the system employs a U-Net model to perform semantic segmentation of 14 anatomical structures in the knee joint, including bones, cartilage, ligaments, and menisci. Next, a ResNet50 model detects tears within the segmented meniscus region. The segmentation network integrates U-Net with transfer learning by replacing the original encoder with VGG16 and incorporating an ASPP module to enhance segmentation performance. Experimental results on 94 knee MRI datasets demonstrate high MIoU for knee joint segmentation and excellent diagnostic accuracy for meniscus tear detection, highlighting the potential and applicability of deep learning in knee joint diagnostics.