Crns: CLIP-driven referring nuclei segmentation

Nuclei segmentation models significantly improve the efficiency of nuclei analysis. Current deep learning models for nuclei segmentation can be divided into single-path and multi-path approaches. Single-path algorithms often underestimate the importance of edge supervision, while multi-path algorithms typically share layers but leading to potential negative impacts on feature extraction due to gradient updates during backpropagation. To address these challenges, we introduced a novel CLIP-Driven Referring model. Specifically, we designed a Class Guidance block that guides the model in distinguishing and aggregating different features by computing the similarity between images and text. We also introduced a Deformable Feature Attention block in the image branch to enhance local modeling abilities. We analyzed DICE, AJI and PQ metrics improvements through cross-dataset validation. Our model achieved increases of 4.14%, 5.69% and 9.06%, respectively, on the CPM when training with MoNuSeg, and 2.16%, 3.85% and 2.86%, respectively, on the MoNuSeg when training with CPM.