Histo-Diffusion: A Diffusion Super-Resolution Method for Digital Pathology with Comprehensive Quality Assessment

TL;DR

Histo-Diffusion introduces a diffusion-based approach for histopathology super-resolution that couples a SwinIR-based restoration stage with a controllable diffusion module conditioned on histology priors. The method uses a dual-stage architecture and a histology-tailored evaluation pipeline, combining full-reference and no-reference metrics to accurately assess texture, intensity, and perceptual realism. Empirical results across PRAD, LUAD, and GBM demonstrate superior perceptual quality and robustness against GAN-based methods, along with multi-scale and WSI-level SR capabilities and demonstrated downstream task compatibility. Together with a curated histopathology IQA dataset and CLIP-IQA-based quality assessments, the work provides a comprehensive framework for reliable SR in digital pathology with practical diagnostic relevance.

Abstract

Digital pathology has advanced significantly over the last decade, with Whole Slide Images (WSIs) encompassing vast amounts of data essential for accurate disease diagnosis. High-resolution WSIs are essential for precise diagnosis but technical limitations in scanning equipment and variablity in slide preparation can hinder obtaining these images. Super-resolution techniques can enhance low-resolution images; while Generative Adversarial Networks (GANs) have been effective in natural image super-resolution tasks, they often struggle with histopathology due to overfitting and mode collapse. Traditional evaluation metrics fall short in assessing the complex characteristics of histopathology images, necessitating robust histology-specific evaluation methods. We introduce Histo-Diffusion, a novel diffusion-based method specially designed for generating and evaluating super-resolution images in digital pathology. It includes a restoration module for histopathology prior and a controllable diffusion module for generating high-quality images. We have curated two histopathology datasets and proposed a comprehensive evaluation strategy which incorporates both full-reference and no-reference metrics to thoroughly assess the quality of digital pathology images. Comparative analyses on multiple datasets with state-of-the-art methods reveal that Histo-Diffusion outperforms GANs. Our method offers a versatile solution for histopathology image super-resolution, capable of handling multi-resolution generation from varied input sizes, providing valuable support in diagnostic processes.

Figures (15)

• Figure 1: Comparisons of state-of-the-art methods and Histo-Diffusion.
• Figure 2: Evaluation examples using PSNR, SSIM, and LPIPS in the field of digital pathology image super-resolution. Generative super-resolution methods produce images with sharper details and closer resemblance to high-resolution ground truths compared to bicubic interpolation. However, bicubic images often achieve higher PSNR, SSIM scores. Despite higher scores, bicubic images appear blurrier to human observers, indicating a disconnect between these metrics and human perception of image quality.
• Figure 3: Generated super-resolution images using GAN-based methods. These methods struggle to preserve stain color in histopathology images. The zoomed-in regions in the right corner of the high-resolution ground truth image, highlights that the GAN-generated super-resolution images exhibit over-smoothing and lack critical texture information within the green box, which are very crucial to accurate diagnosis.
• Figure 4: Dual-Stage Diffusion-Based Image Super-Resolution Model. It includes a restoration module that generates restored images as histopathology priors for a controllable diffusion module. The restored and noisy latent images are combined to work as the input for the controllable diffusion module for super-resolution image generation.
• Figure 5: Restored images with corresponding decoded control images using VAE with condition latent $c_{RM}$