D-DDPM: Deep Denoising Diffusion Probabilistic Models for Lesion Segmentation and Data Generation in Ultrasound Imaging

Abstract

The Denoising Diffusion Probabilistic Model (DDPM) has gained significant attention for its powerful image generation and segmentation capabilities, particularly in biomedical applications where accuracy is critical. In breast cancer detection, ultrasound imaging is widely used due to its safety, affordability, and non-ionizing nature. However, the inherent challenges of ultrasound data, such as noise and artifacts, make accurate tumor segmentation difficult, often leading to misdiagnosis. We propose a novel Deep Denoising Probabilistic Diffusion Model (D-DDPM) designed to enhance tumor segmentation in breast ultrasound images to address these limitations. Our model incorporates a nested U-Net architecture with Residual U-blocks (RSU), significantly improving feature learning and segmentation precision. In addition to performing segmentation, D-DDPM generates synthetic data, augmenting existing real datasets to improve data size with a diverse range of high-quality samples. We validated D-DDPM on several breast ultrasound datasets, comparing its performance to state-of-the-art methods. The proposed D-DDPM achieves a Dice score improvement of 2.26%, 4.24%, and 5% over the runner-up model, demonstrating superior performance on all BUS datasets. Both qualitative and quantitative results demonstrate the ability of D-DDPM to deliver more accurate and reliable segmentation results, offering promising potential to improve clinical decision-making in cancer diagnosis.

Keywords

• Gaussian noise
• diffusion
• denoising
• sampling
• segmentation
• ReSidual U-blocks