SYNTHETIC HISTOPATHOLOGY LYMPHOCYTE IMAGES AUGMENTATION USING CONDITIONAL GENERATIVE ADVERSARIAL NETWORKS

Abstract

Histopathology image analysis is widely used and is essential for diagnosis and cancer grading, including colorectal cancer. However, due to a lack of availability and labor-intensive annotation procedures, getting a significant and varied collection of histopathology images for training machine learning models continues to be difficult. To solve this problem, we suggest a Latent-to-Image method that creates synthetic colorectal histopathology images using Conditional Generative Adversarial Networks (cGANs). In this article, we investigate the use of cGANs specifically for generating images for lymphocytes tissue, using seven different tissue classes for training. The results show that the generated synthetic images are indistinguishable from the real histopathological images, capturing the distinctive textural and structural characteristics of the tissue. We show that our method generates high quality images with a Fréchet Inception Distance of 21.2. The generated images were also assessed by four pathologists and no significant difference between the real and generated images were found.