Generation of Lung Cytological Images Using Conditional Diffusion Model

Abstract

Cytology imposes a significant burden on practitioners, and image classification techniques have been developed to alleviate this burden. One approach to improving the performance of image classification models is data augmentation using generative AI. However, conventional generative techniques often introduce model-specific artifacts, making it difficult to generate realistic images. In this study, we propose a novel conditional diffusion model to generate high-quality benign and malignant lung cytology images for use in image classification models. We conducted a subjective evaluation by cytology experts and a quantitative evaluation based on image similarity, comparing the images generated by our proposed method with those generated by conventional techniques. The results demonstrated that our proposed method outperformed conventional approaches. In conclusion, the proposed method may be effective in generating realistic cytology images.