Noise-Consistent Siamese-Diffusion for Medical Image Synthesis and Segmentation

Deep learning has revolutionized medical image segmentation, but its full potential is limited by the scarcity of annotated datasets. Diffusion models are used to generate synthetic image-mask pairs to expand these datasets, yet they also face the same data scarcity issues they aim to address. Traditional mask-only models often produce low-fidelity images due to insufficient generation of morphological characteristics, which can catastrophically undermine the reliability of segmentation models. To enhance morphological fidelity, we propose the Siamese-Diffusion model, which incorporates both image and mask prior controls during training and switches to mask-only guidance during sampling to preserve diversity and scalability. This model, comprising both Mask-Diffusion and Image-Diffusion, ensures high morphological fidelity by introducing a Noise Consistency Loss between the two diffusion processes, guiding the convergence trajectory of Mask-Diffusion toward higher-fidelity local minima in the parameter space. Extensive experiments validate the superiority of our method: with Siamese-Diffusion, SANet achieves mDice and mIoU improvements of 3.6% and 4.4% on the Polyps dataset, while UNet shows mDice and mIoU improvements of 1.52% and 1.64% on the ISIC2018 dataset. Code will be released.