Diffusion model for breast cancer segmentation
Xue Wang, Lianggui Tang, Hongda Mou, Song Peng, Xiuling Zhu, Xuan Lai · 2023
Breast cancer has become one of the leading causes of death among women, and in recent years, the incidence of breast cancer in China has been steadily increasing. However, breast cancer is a chronic disease with no obvious early symptoms, and once abnormalities are discovered, they are often in the middle or late stages. Clinical data shows that early detection and treatment can greatly improve the survival rate of patients. Breast cancer ultrasound imaging technology has gradually become the mainstream method of breast cancer diagnosis due to its low cost and non-invasive safety features, and the emergence of computer-aided diagnosis systems can better assist doctors in reading and diagnosing images, reducing misdiagnosis and missed diagnosis rates. Based on this, in order to improve the accuracy of existing breast cancer instance segmentation methods, we proposes a diffusion-model-based breast cancer instance segmentation method, Diff-bcSeg, which defines breast cancer instance segmentation as a denoising process from noise to filters. The trained model can reverse the noisy ground truth without any inductive bias from RPN, using randomly generated filters as input and output masks for one or more denoising steps. The proposed method is validated on the Breast Ultrasound Images (BUSI) dataset, and experimental results demonstrate that Diff-bcSeg has better detection and segmentation performance for breast cancer instance segmentation compared to current segmentation models.