Conditional Dual-Branch Diffusion Network for Cervical Cell Nucleus Segmentation in Cancer Diagnostics

Cervical cancer is one of the most common malignancies among women worldwide and carries a high mortality rate, posing an even greater threat in developing countries where medical resources are limited. Automatic analysis of cervical cytology images is an important means of improving screening efficiency and reducing incidence. However, the diversity of staining methods and the complex morphology of cells make it difficult for traditional hand-crafted-feature–based segmentation methods to handle real-world challenges such as adhesion and overlap. Although convolutional neural network (CNN)–based segmentation models have improved performance to a certain extent, they are inherently discriminative and struggle to model the joint generative relationship between images and masks, which limits structural reconstruction accuracy. To overcome these limitations, we propose a dual-branch segmentation framework based on diffusion models that combines generative modeling with discriminative semantic guidance. The framework consists of a diffusion backbone that learns the image–mask generative relationship and a conditional branch that supplies semantic priors for guidance while also producing a coarse segmentation result for auxiliary supervision. A dual-stream feature fusion module is introduced to strengthen information exchange between the two branches. Experiments on multiple cervical cell datasets show that the proposed method surpasses existing mainstream approaches in both accuracy and robustness for nucleus segmentation.