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

Area of Science:

• Computational Biology
• Medical Imaging
• Artificial Intelligence

Background:

• Histopathological image acquisition is challenging, leading to limited diversity in generated cell morphology.
• Existing methods struggle to capture the full spectrum of cell shape variations.

Purpose of the Study:

• To develop a novel diffusion generation model for creating diverse and detailed cell morphology images.
• To address the lack of diversity in current cell morphology generation techniques.

Main Methods:

• Proposed the first point diffusion-based generative model for cell morphology.
• Incorporated a class mask self-attention module to control generated cell types.
• Utilized gradual information updating to enhance realism and diversity during image generation.

Main Results:

• Achieved superior performance on the Lizard public dataset compared to existing methods.
• Demonstrated a 43.17% improvement in Fréchet Inception Distance (FID) and a 46.24% enhancement in Inception Score (IS) over the NASDM network.
• Successfully generated diverse and high-quality cell morphology images.

Conclusions:

• The proposed diffusion model, integrating point diffusion and class mask self-attention, is a pioneering approach.
• The model effectively generates diverse datasets while preserving high image quality.
• Experimental results confirm the model's excellent performance in cell morphology generation.