Architecture of Multiple Convolutional Neural Networks to Construct a Subject-Specific Knee Model for Estimating Local Specific Absorption Rate

Abstract

Electromagnetic simulation is a credible way to estimate the local specific absorption rate (SAR), which is a key consideration in high-field magnetic resonance imaging of the knee joint. To construct a subject-specific knee model, which is critical for SAR simulation, we proposed an architecture comprising multiple convolutional neural networks. Knee tissues were segmented by three U-Nets. Each network was responsible for the segmentation of two tissues that have relatively similar volumes and distinct intensity distributions (muscle and fat, cancellous and cortical bone, and cartilage and meniscus). Additionally, a weighted loss function was used to further alleviate the effect of class imbalance of the segmented tissues. The outputs of these three networks were merged and morphological filtering was used as the post-processing to eliminate holes and isolated voxels. This method was compared with three other segmentation methods. Good segmentation performance was demonstrated on the test set, and the proposed method was found to be superior to the other methods according to several quantitative measures. Meanwhile, local SAR in a 3 T coil using models constructed with the proposed method, manual delineation, and the comparison methods were also evaluated on the test set. On the whole, the maximum values of SAR_10g of the models constructed by the proposed method were closer to the results of manual delineation. Overall, the proposed method exhibits promising potential for precisely constructing knee models for SAR simulation.