Segmentation of the knee tissues using U-Net neural network based on T1- and T2-weighted MR images

Abstract

Objective: To automatically and precisely segment knee tissues for aiding diagnosis or constructing model, a segmentation method using a convolutional neural network which was based on the data set of T1- and T2-weighted MR (magnetic resonance) images was proposed in this paper. Materials and methods: The data set consisted of training set, validation set, and test set, and each of them had image sets of 14, 4, and 12 volunteers, respectively. Each knee image set integrated T1- and T2- weighted MR images in the sagittal plane from a low-field scanner. A fully convolutional network, U-Net, was used to segmented knee tissues, which was composed of cartilage, meniscus, effusion, bone, muscle, and fat. For U-net, convolution layer number was 13, kernel size was 7x7, and cross entropy function was used as the loss function. To eliminate isolated pixels, output images were processed using morphological filtering. This method was compared with those methods that only used T1- or T2-weighted images by several quantitative measures. The manual delineation results were used as the ground truth. Results: Good segmentation performance was demonstrated on the test set. The quantitative measures of most tissues of the proposed method were found to be superior to those of other methods mentioned above. Conclusions: The proposed method adopted an advanced neural network and implemented a comprehensive use of information contained in T1- and T2-weighted images. Therefore, it exhibited promising potential for automatic and precise segmentation of knee tissues.