Prior-guided Diffusion Model for Cell-level Segmentation in Quantitative Phase Imaging

Quantitative phase imaging (QPI) produces high-contrast, label-free images of tissue and cell samples without dyes. Accurate cell-level segmentation in QPI is essential for various biomedical applications. Diffusion models (DM) have shown strong performance in segmentation, surpassing U-Net and DeepLabv3. Unlike deterministic methods, DM employs multiple random samplings from a Gaussian distribution to generate different predictions and uses ensemble learning for improved accuracy. However, the lack of content information in the starting noise and the need for multiple samplings limit DM's efficiency and accuracy. Leveraging the high-contrast content in QPI images, we introduce a prior-guided DM-based segmentation method. We replace multiple random starting noises with content-informed noise using a single sampling, significantly enhancing both speed and accuracy. Since we do not change the DM training process, the prior-informed noise can be integrated into various DM-based frameworks. Experiments on two QPI datasets confirm the effectiveness of our method.