Automated Deep Learning Segmentation of Human Knee Cartilage from 3T MRI with Boundary Information

Knee Osteoarthritis (OA) is serious and prevalent today. Image segmentation of high-resolution MRI scans measuring cartilage volume and thickness is useful to track knee OA progression in the early stages and avoid joint replacement. In this work, we developed a cheap and efficient automated technique based on U-Net for knee cartilage segmentation, paying more attention to boundary information. Our model outperforms many existing models for segmentation of the femoral cartilage and performs as well as other techniques for other cartilage compartments. The boundary loss appears to improve cartilage segmentation for the edge slices with smaller cartilage volume.