Abstract

The field of histopathology faces significant challenges due to the limited availability of data, which is often not publicly accessible due to privacy issues. The scarcity of high-quality publicly available datasets hampers the development and training of effective deep learning models. Generative Adversarial Networks (GANs) have previously attempted to address these issues by creating synthetic data but suffer mode collapse, which reduces their effectiveness and reliability. This study explores Diffusion Generative Models (DGMs) as a unique and robust alternative for generating synthetic pathology images.
The primary objective of this study is to compare various Diffusion Generative Models (DGMs) and methods in medical imaging. Specifically, we examine the Denoising Diffusion Probabilistic Model (DDPM) and the Latent Diffusion Model (LDM), along with other generative sampling choices. Both models demonstrated the ability to generate realistic histopathological images. We also investigated DGMs from a unique perspective by generating various patch sizes, demonstrating that DGMs effectively learn patch resolution.
We analyzed the impact of DGMs on different magnifications in the KGH dataset, focusing on image patches of 224x224 and 336x336 pixels. Larger patches (336x336) showed better performance, with real data achieving 94.06% accuracy and generated data 92.44%. However, combining real and generated data slightly reduced accuracy to 90.76%. For 224x224 patches, real data achieved 89.95%, generated data 88.62%, and combined data improved to 90.75%. These results indicate that synthetic data enhances model performance, particularly with larger image patches.
In computational pathology, generative models can enhance data sharing and augmentation, improving the accuracy of deep learning classifiers, and assisting in the cancer diagnosis workflow, thereby advancing digital pathology. Our research findings confirm this and set the stage for future developments in the pathology workflow.