Generation and Validation of Synthetic Mammography Images Using Diffusion and Blending Approaches

Fatima Alshihri, Aqeelah J Makki, Fay A. Al Shammari, Abdullah A Alsheeha, Shahad M Albati, Pedro Jose Moreno Mengibar, Mona F Almehaid, Salman Albshan, Hend Samir Ibrahim, Manal Ahmed ElRefaei, Fatma Eliraqi, Yousef Alkathiri, Ammar Almoalem, Hamad Al-msalam

Keywords: Stable Diffusion 3.5, Inpainting, Poission blending, Mamography, Otsu

Abstract: Breast cancer is among the most prevalent diseases that affect women and remains a significant global health concern. Because medical imaging is highly confidential, publicly accessible mammography datasets are limited. This shortage has forced us to explore alternative ways of obtaining reliable training data. Generative models serve as an alternate data augmentation method to mitigate the data scarcity issue encountered in medical imaging. Diffusion models have garnered significant attention due to their novel generation methodology, the superior quality of the produced images, and their comparatively simpler training process. In this paper, we explore the use of three different techniques: Stable Diffusion 3.5, Poisson blending, and Stable Diffusion Inpainting for generating synthetic mammography images. Using these methods, we created thousands of images that span multiple categories of BI-RADS, breast laterality, and standard mammographic views (MLO and CC). To evaluate the generated images, we invited seven radiologists to review them using a dedicated assessment tool. Each radiologist was asked not only to decide whether an image was real or synthetic but also to assign a BI-RADS category as would be in routine practice. What stood out in the results was the amount of confusion: about 37\% of the synthetic images were judged to be real, and approximately 30\% of the authentic images were misidentified. The difficulty experienced by Radiologists in distinguishing between the two indicates that synthetic images are nearing the visual complexity of real mammograms. These discoveries indicate a distinct opportunity to utilize synthetic mammograms in research, particularly in contexts where access to extensive, diverse clinical datasets is constrained.

Primary Subject Area: Generative Models

Secondary Subject Area: Image Synthesis

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 338