Fully Automatic Segmentation of Knee Joint Anatomy and Lesions using Clinical MR Images

Motivation: Osteoarthritis affects multiple tissues in the knee joint. However, there is a lack of an efficient method for automatic segmentation of tissues and lesions using a single clinical MRI sequence. Goal(s): To provide a solution for automatic segmentation of femur and tibia bone and cartilage, plus bone marrow edema-like lesions (BMEL) using IW-TSE images only. Approach: We trained a multi-label segmentation model in a supervised manner, employing pre- and post-processing steps to improve its robustness and stability. Results: We find that a lightweight convolutional neural network can be trained to segment the five regions with a combined Dice similarity coefficient (DSC) of 0.87. Impact: We provide an efficient and consistent solution for the segmentation of knee joint anatomy and lesions, enabling large-scale downstream analyses without incurring large costs for manual annotations.