Cross-Modality Fourier Feature for Medical Image Synthesis.

In this paper, we propose a cross-modality fourier feature (CMFF) method via frequency selection, which learns the rational anatomical structure for targeting medical modality images. Unlike existing works seeking pixel-wise intensity discrepancy likely misleading to bias anatomical structures, our approach strongly holds the medical prior that different modality MRI images should share the same anatomical structure. To achieve this, our method instead learns to convert MRI images to the auxiliary frequency domain. Moreover, we adopt the Shapley value to quantify the contribution of each frequency, with respect to the structure for MRI image pairs of different modalities. Thus our approach learns to refine the anatomical structure generated by the target modality iteratively. Extensive experimental results on the BraTS dataset show that our model surpasses the performance compared to SOTAs.