Knee Model Construction Based on MR images Using U-Net and Conditional GAN

Abstract

Purpose: Electromagnetic (EM) simulation with a knee model is the main method to calculate local SAR of a knee joint in high-field MRI. Since the length of knee models used for simulation affects the calculation results of SAR value, a knee model construction method based on MR images using U-Net and conditional generative adversarial network (CGAN) was proposed in order to construct approximate knee models and thus obtain more accurate local SAR results. Materials and Methods: Knee tissues were simplified based on the “muscle-fat-bone” simplification, all tissues except fat and bone were classified as muscle. The U-Net was used to classify the three tissues of the original sagittal image with a field of view (FOV) of 150 mm × 150 mm. Then the CGAN was used to generate these tissues outside both ends of the knee joint and extend the FOV to 230 mm × 150 mm. Therefore a knee model with length of 230 mm instead of 150 mm was constructed. Result and discussions: EM simulation (length of coil model was 180mm) was performed and local SAR was calculated for the models obtained from the proposed method and comparison methods, and their relative errors of the maximum SAR <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10g</inf> with models by manual delineation (full tissues) were calculated. For the proposed method, the mean and standard deviation of the relative errors were 0.0814 and 0.0531, respectively. Conclusion: The proposed method based on the “muscle-fat-bone” simplification and an architecture integrating U-Net and CGAN is capable of generating a knee model, through which a relative accurate local SAR can be obtained.