Cytofusion: A Latent Diffusion-Based Framework for Cytology Classification

With advancements in screening and storage techniques, cytology classification has become crucial for accurate and effective clinical decision-making. Recently, computer-aided diagnosis has emerged as a vibrant research area, with many automated classification systems achieving outstanding performance in medical image analysis. However, real-world cytology datasets often suffer from data imbalance and scarcity, requiring substantial human and material resources to curate high-quality datasets, which in turn limits the performance of deep learning models. To address these challenges, we propose cytology diffusion model - Cytofusion, a data augmentation framework based on a latent diffusion model, marking the first application of diffusion models to cytology image generation. Cytofusion employs advanced architectural enhancements and optimized loss functions to refine feature representation, enable multi-scale fusion, and preserve cellular structures, ensuring high morphological fidelity and structural consistency in synthesized cytology images. Extensive experiments conducted on three public cytology datasets demonstrate that Cytofusion significantly outperforms existing augmentation techniques, providing a robust and effective solution for advancing cytology classification.