A Deep Learning-Based Method for Knee Articular Cartilage Segmentation in MRI Images

Abstract

MRI image segmentation of knee articular cartilage is important for the early diagnosis of osteoarthritis. Unfortunately, manual segmentation of articular cartilage consumes a lot of time, and is affected by subjective factors and prone to errors leading to unstable accuracy. In this study, we propose a deep learning-based method for automatically segmenting knee articular cartilage in MRI images, wherein a spatial attention mechanism is dedicatedly added into the used UNet structure for enhancing comprehensive performance. A total of 80 persons are included in the experimental dataset, wherein 18 slices per knee joint, 2520 images of 70 persons in the training set, and 360 images of 10 persons in the test set. The proposed method achieves a good performance, and the final Dice similarity coefficient (DSC) calculated is 88.7%• After a successful segmentation, it is expected that the thickness and density of the knee articular cartilage can be evaluated effectively and efficiently, which would be helpful for the early diagnosis osteoarthritis.