Cell generation with label evolution diffusion and class mask self-attention.

Purpose: Due to the relative difficulty in acquiring histopathological images, the generated cell morphology often presents a fixed pattern and lacks diversity. To this end, we propose the first diffusion generation model based on point diffusion, which can capture the changes and diversity of cell morphology in more detail.

Methods: By gradually updating the information of cell morphology during the generation process, we can effectively guide the diffusion model to generate more diverse and realistic cell images. In addition, we introduce a class mask self-attention module to constrain the cell types generated by the diffusion model.

Results: We conducted experiments on the public dataset Lizard, and comparative analysis with previous image generation methods showed that our method has excellent performance. Compared with the latest NASDM network, our method achieves a 43.17% improvement in FID and a 46.24% enhancement in IS.

Conclusions: We proposed a first-of-its-kind diffusion model that combines point diffusion and class mask self-attention mechanisms. The model can effectively generate diverse data while maintaining the high quality of generated images and performs well.