Meniscus segmentation and detection of meniscus tears in MR images

Medical data contains valuable information suitable for diagnosis of diseases. However, the size and complexity of data sets make it difficult to classify data. This provides that automatic detection systems become common on medical data. One of these data is the meniscus structures of the knee joint which is the subject of this thesis study. Meniscus tears are one of the common knee disorders, especially in sports and elderly people. Therefore, the right time to put the right diagnosis ensures that the various disorders that may occur in the knee can be avoided, such as osteoarthritis. This study suggests new computer-based and fully automated approaches to support radiologists (i) to segment menisci and to detect tears, (ii) to provide early diagnosis and treatment, and (iii) to reduce errors caused by MR reader differences. In our studies, we used MR images in the water-selective excitation) double echo in the steady state (weDess) standard obtained in the sagittal plane provided by the Osteoarthritis Initiative (OAI) for these purposes. Two different and comprehensive studies were carried out within the scope of the thesis. In the first study, extreme learning machine (ELM) and random forest (RF) methods were used for model learning (regression) and histogram of oriented gradients (HOG) and local binary patterns (LBP) for feature extraction. First of all, there are the most compact rectangular windows that limit meniscus. After that, the meniscus boundaries are obtained by morphological processes. Then, similarities between predicted boundaries and ground truth boundaries are measured and compared with each other. The highest meniscus segmentation success achieved with the Dice similarity metric was 82.73%. In the second study, a new and different approach to the previous method was proposed for the segmentation of menisci and the automatic classification of meniscal tears. This study consists of three basic stages: preprocessing, segmentation and classification. In the preprocessing step, the acquisition of the windows where the menisci are located was performed from the MR slices. The meniscus structures were segmented by fuzzy c-means (FCM), spatial fuzzy c-means (sFCM) and improved spatial fuzzy c-means (isFCM) clustering methods at the segmentation step. The k-nearest neighbor (kNN), extreme learning machine (ELM) and support vector machines (SVM) classifiers were used to classify segmented images and to detect meniscus tears. The method first decides whether there are tears on menisci; if this is the case then, determines the place (anterior horn, meniscus body, posterior horn) and the type of tears (horizontal, vertical, etc.) with high success rates within 3-4 minutes. The proposed system realizes classification of meniscus tear types by a success rate of 84.97%, which has not been done before in the literature. The computer aided diagnostic systems (CAD) proposed in the study can be used by radiologists as a decision support system in meniscus segmentation and diagnosis of meniscus tears with these aspects. Keywords: Segmentation, Knee joint, Meniscus, Medical Images, Computer aided diagnosis, Meniscus tears, magnetic resonance imaging.