This paper addresses the issue that segmentation of mass lesions in digital breast tomosynthesis (DBT) images is very important for early diagnosis of breast cancer, but mass lesions are easily obscured by dense breast tissue, making manual annotation difficult and time-consuming, and thus lacking annotation data for model training. Existing diffusion model-based data augmentation methods have limitations in the quality of generated images and their usability due to the difficulty in learning lesion region features and the difficulty in generating image and annotation pairs. In this paper, we propose a method to generate paired images by training an additional diffusion guider on a conditional diffusion model without external conditions. Through experiments, we generate DBT slice and mass lesion mask pairs and integrate them into the supervised learning process of the mass lesion segmentation task, thereby verifying the performance improvement. The proposed method improves the generation quality without external conditions and alleviates the problem of insufficient annotation data, thereby contributing to improving the performance of subtasks.
