Conditional image synthesis for improved segmentation of glomeruli in renal histopathological images

In a context of limited data availability, we consider the supervised segmentation of glomerular structures in patches of renal histopathological whole slide images. These structures are complex, include multiple substructures, and exhibit great variability in their shape, making their robust segmentation challenging. In this context, using appropriate data augmentation techniques is crucial to obtain more robust results. We investigate data augmentation based on random spatial deformations and conditional image synthesis for the training of a U-Net model. We rely on a SPADE model to perform the synthesis, using label maps built from the real patches available for training as input. Synthesis from ground truth masks only results in noisy patches, where substructures are absent, whereas additional structure information yield more realistic patches. We show that the best improvements of the segmentation performances are obtained by mixing real patches with synthetic patches generated from ground truth masks only, which yields an increase of up to 0.76 of average dice score w.r.t. augmentation based on spatial deformations only. We conclude that, using conditional image synthesis, patches synthesized with no additional structure information better contribute to the robustness of glomeruli segmentation than patches synthesized with structure information extracted from available real patches.