Denoising Diffusion Probabilistic Models for Generating Tissue Type Breast Image Dataset

Various supervised deep learning techniques have been used to enhance microwave imaging for breast cancer detection, but such techniques require large amounts of labelled training data which remains a significant challenge. In this research, we addressed the scarcity of realistic numerical permittivity breast phantoms by developing a generative model. This model uses high-level annotations, including breast types defined by BI-RADS categories, to produce a large, labeled tissue-type breast image dataset. Our prototype model integrates a U-Net with a denoising diffusion probabilistic model to generate tissue-specific 2D sagittal cross-section images of the segmented breast. Initial results demonstrate the model’s proficiency in generating realistic images that closely resemble segmented MRI scans, with quantitative assessments through Fréchet Inception Distance (FID) scores confirming their authenticity. The prototype model will be extended to 3D in future work.