A Rapid Method for Constructing an Approximate Knee Model for Local SAR Estimation Based on a Convolutional Neural Network

Abstract

Purpose: The local specific absorption rate (SAR) of tissues is a key consideration in high-field MRI of knee joint. To automatically and rapidly construct individual’s knee model for SAR simulation and estimation, a segmentation method for MR images using U-Net network and tissue simplification is proposed in this paper. Materials and methods: Based on a data set of low-field MR Tl-weighted images in the axial plane, the knee tissues were simplified and classified as muscle, fat, and bone for labeling and segmentation. Single U-net was adopted to segmented images, the number of convolutional layers was 13, the convolution kernel size was 7x7, and the mini batch was 16. The data set consisted of 30 volunteers’ knee joint images (in the axial plane): 14, 4, and 12 for the training, validation, and test set respectively. To overcome the limitation of imaging slice number and make the SAR simulation result more accurate, the constructed knee model was extended in the axial direction using slice extrapolating. Based on the test set, electromagnetic simulation in a 3T coil was carried out and local SAR was evaluated using the models constructed with the proposed method and manual delineation. Result and discussions: The two constructed models had similar SARlOg distribution, the locations of hot spots were basically unchanged, and the mean and standard deviation of the errors between their maximum SARlOg values were relatively small. Conclusions: The method by using single U-Net networks, tissue simplification and model extending shows promising potential for constructing knee model on which approximate local SAR could be estimated.