Knee Joint Model Construction and Local Specific Absorption Rate Estimation Based on Generative Adversarial Networks

Abstract

The local specific absorption rate (SAR) is an essential metric when assessing the safety of high-field magnetic resonance imaging (MRI). Usually, the subject-specific model is created from segmented magnetic resonance (MR) images, followed by SAR estimates. However, the length of knee joint model affects the accuracy of local SAR estimation in simulation. To address the issue, this paper proposed a knee joint MR image segmentation and field-of-view extension method based on conditional generative adversarial nets (CGAN). The image of knee joint was classified into images of three tissues, namely, muscle, fat, and bone. And pixel semantic segmentation was performed using CGAN with attention mechanism to improve accuracy. The method also generated extension areas at both ends of the image along the head-foot direction to construct a longer model. The knee joint models by using the proposed method and various comparison methods were electromagnetically simulated, and the relative error of their local SAR compared to the manually annotated model was calculated. Results verified that the proposed method can achieve a relatively accurate estimation of local SAR in the knee joint.

Keywords

• magnetic resonance imaging
• local specific absorption rate
• conditional generative adversarial nets
• knee joint
• image segmentation