Knee Model Construction Using Deep Neural Networks with Boundary Information for Local SAR Estimation

The local specific absorption rate (SAR) is a key safety indicator in high-field MRI. Constructing a specific model for each patient is important for accurate estimation of local SAR. The aim of this study is to construct subject-specific knee models based on low-field images for realizing accurate local SAR estimation in high-field MRI systems (3T and 1.5T). The proposed method used two U-Net networks for tissue segmentation of knee joint and the classification results of the two networks were merged to generate the final models. Muscle has high dielectric properties and large volume, which have an important influence on the electromagnetic field distribution. To improve the accuracy of muscle segmentation, a U-Net making use of boundary information was used to segment muscle alone to overcome the problem of inhomogeneous intensity at the edge of the muscle region. Other tissues were segmented by another U-Net, which used a weighted loss function to mitigate the adverse influence of class imbalances between tissues. The proposed method was compared with other methods using manual delineation as the standard. Its muscle segmentation performance was better than that of the comparison methods. On the other hand, local SAR in 3T using models constructed by these methods was also evaluated through electromagnetic simulation separately. It was shown that the maximum SAR_10g of the models constructed by the proposed method was much closer to that of manual delineation on the whole. These results validated the availability of the proposed method.