High-Resolution Histopathology Whole Slide Image Generation Using Wavelet Diffusion Model

Recent advances in deep learning (DL) have significantly improved computational pathology, particularly for analyzing tissue images in diagnostic and prognostic tasks. However, most DL methods depend on large-scale annotated datasets, which are difficult to obtain in the medical field due to the time, cost, and labor involved in annotation. To address this, generative models, especially diffusion models, have been investigated to create synthetic whole-slide images (WSIs). However, generating high-quality histopathology WSIs is challenging due to their gigapixel resolution. While diffusion models produce excellent images, they suffer from slow inference speeds, limiting their practical application. In this work, we propose a Wavelet-Diffusion Model (WDM) that integrates wavelet transforms into diffusion models, enhancing sampling efficiency without compromising image quality. The resulting framework, WSI-WDM, was evaluated on two public datasets: PAIP2019 (liver cancer segmentation) and BCSS (breast cancer semantic segmentation). Our experimental results show that WSI-WDM outperforms state-of-the-art methods in both generation quality and inference speed. Specifically, it achieved Frechet Inception Distance (FID) scores of 268.55 for PAIP2019 and 303.02 for BCSS, with inference times of $300.05\pm 0.84$ ms and $275.17\pm 0.72$ ms, respectively, demonstrating that WSI-WDM provides an efficient, high-quality solution for generating synthetic WSIs.