Scale Feature-Aware Generative Adversarial Network Improve MRI Device Data Imbalance for Healthy Consumption

Abstract

In consumer healthcare, the imbalance between the imaging data of MRI devices creates a significant barrier to the generalization ability of medical image segmentation. We employ generative adversarial networks to address the challenges inherent in the heterogeneity of data distribution among different MRI devices. First, we propose the Scale Feature Awareness Module (SFAM), which can skillfully capture image details at different scales and broader contextual information. Then, we propose the Dynamic Scale Attention Module (DSAM), which aims to combine feature mappings at different scales for different paths dynamically. We can assign different dynamic weights to each scale to enhance the feature representation. Finally, we propose a multi-scale discriminator to guide the generation of aneurysm images with different diameters. Based on the above modules, we designed Scale Feature Aware Generative Adversarial Network (SFAGAN) to generate medical images with the same distribution. It is experimentally demonstrated that SFAGAN improves PSNR, SSIM, and FID values by 0.64, 0.62, and 7.99, respectively, over the SOTA method. In addition, we use the generated data for downstream segmentation tasks to demonstrate the quality of the generated images. Notably, our SFAGAN has significant performance, making its application in medical systems very feasible.