A U-Net Network-Based Rapid Construction of Knee Models for Specific Absorption Rate Estimation

Abstract

The specific absorption rate (SAR) needs to be estimated for safety concerns when performing high-field magnetic resonance imaging (MRI) on knee joint. Electromagnetic simulation can be used to calculate the local SAR on knees if a patient-specific knee model is available. In this work, a method for rapid construction of knee models from low-field magnetic resonance knee images was proposed. A convolutional neural network (CNN) was first used to segment the sagittal T1-weighted spin echo images of knees into fat, bone and muscle. All the pixels other than those from the fat and bone are classified as muscle pixels. The U-Net network, a full CNN with a convolutional layer number of four and adopting a cross entropy function, was used to perform segmentation slice-by-slice. The results from automatic segmentation were compared with those obtained with manual delineation with quantitative measures. Moreover, SAR with a 3 T quadrature birdcage coil was calculated. The simulation results validated the proposed method by showing that and a relatively accurate local SAR estimation could be obtained with the knee models constructed from low-field knee images.

Keywords

• high-field magnetic resonance imaging (mri)
• specific absorption rate (sar)
• knee joint
• u-net network
• image segmentation