Fig. 2: Qualitative results of the CTA heart experiment.
From: Mining multi-center heterogeneous medical data with distributed synthetic learning
a Four image generation examples for the GAN-based methods. Each row has the multi-component mask image of the heart (Label), which is the input of the image generation, the corresponding real CTA image (Real), and the synthetic images generated by different methods (FLGAN, FedMed-GAN, AsynDGAN, DSL). DSL generates images with higher scores in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). b The Dist-FID and FID score curves over the training epochs. FID is calculated using real data from all data centers. The red circles indicate the best epoch for each method, while the arrows show the consistency of FID and Dist-FID scores. c Three examples of segmentation results for different methods vs the ground truth Label. The segmentation model learned from DSL’s synthetic data obtains more accurate results than the other methods (Real-WHS, Real-CAT08, Real-ASOCA, FLGAN, FedMed-GAN, AsynDGAN) and is comparable to centralized learning (Real-All).
