T-CADiff: Conditional guidance based diffusion model for medical image segmentation

Abstract

Diffusion models receive particular attention in the field of medical image segmentation due to their excellent quality and good image diversity. However, there are still some urgent issues remaining. Diffusion models used for image segmentation require the original image to guide generation, and the semantic features of the original images are often misaligned with the features of noisy segmentation masks, affecting the realism of the generated images. Meanwhile, most existing studies based on diffusion models still rely on the classic UNet structure, which does not fully utilize global information. To address the above issues, the model includes a condition guidance transformer (CGT) module to fuse features of the original images and noisy segmentation masks and learn global information. In addition to noise prediction, the real mask is also predicted and subsequently fed into a discriminator to increase the accuracy of predicted noise and the realism of the segmentation images. On the ISIC 2016 skin lesion segmentation dataset, our model achieved 11.04, 85.35%, 91.51%, 90.93%, and 95.63% on the HD95, IoU, Dice, sensitivity, and accuracy, respectively. It can be demonstrated that the proposed model significantly enhances segmentation accuracy, outperforming the state-of-the-art models.