A new computer-based approach for fully automated segmentation of knee meniscus from magnetic resonance images

Menisci are tissues that enable mobility and absorb excess loads on the knee. Problems in meniscus can trigger the disorder of osteoarthritis (OA). OA is one of the most common causes of disability, especially among young athlethes and elderly people. Therefore, the early diagnosis and treatment of abnormalities that occur in the meniscus are of significant importance. This study proposes a new computer-based and fully automated approach to support radiologists by: (i) the segmentation of medial menisci, (ii) enabling early diagnosis and treatment, and (iii) reducing the errors caused by MR intra-reader variability. In this study, 88 different MR images provided by the Osteoarthritis Initiative (OAI) are used. The histogram of oriented gradients (HOG) and local binary patterns (LBP) methods are used for feature extraction from these MR images along with the extreme learning machine (ELM) and random forests (RF) methods which are used for model learning (regression). As the first step of the pipeline, the most compact rectangular patches bounding the menisci are located. After this, meniscus boundaries are revealed by the morphological processes. Then, the similarities between these boundaries and the ground truth images are measured and compared with each other. The highest score is acquired with Dice similarity measurement with a success rate of 82%. A successful segmentation is performed on the diseased knee MR images. The proposed approach can be implemented as a decision support system for radiologists, while the segmented menisci can be used in classification of meniscal tear in future studies.