Breast Tumor Image Synthesis based on Diffusion Probabilistic Model

Ultrasonography plays a crucial role in the early detection of breast cancer due to its real-time, safe, and cost-effective nature. However, the accuracy of ultrasound-based breast cancer diagnosis is highly dependent on the operator's expertise. Recent advancements in deep neural networks have led to the development of learning-based approaches aimed at automating breast tumor detection and segmentation, thereby supporting clinicians in diagnosis. Learning-based approaches require extensive B-mode image datasets with corresponding labels for effective training. However, configuring such datasets demands substantial time from expert radiologists, inherently constraining the quantity and diversity of available data. Furthermore, ultrasound image characteristics vary based on equipment specifications and image processing algorithms, making models trained on limited data prone to domain shift when applied to images obtained under different conditions. In this paper, we present a novel generative model, a content and style-conditioned diffusion probabilistic model, that synthesizes breast ultrasound images, accurately representing a wide range of anatomical and measurement variations. The proposed model is evaluated through qualitative and quantitative assessments and is applied to the downstream task of breast object detection.