Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

URCDM: Ultra-Resolution Image Synthesis in Histopathology

Sarah Cechnicka, James Ball, Matthew Baugh, Hadrien Reynaud, Naomi Simmonds, Andrew P. T. Smith, Catherine Horsfield, Candice Roufosse, Bernhard Kainz

TL;DR

The paper tackles the challenge of generating coherent, ultra-high-resolution histopathology WSIs that preserve hierarchical structure across magnifications, addressing memory and privacy issues of patch-based diffusion. It introduces Ultra-Resolution Cascaded Diffusion Models (URCDMs), a three-stage cascade of diffusion processes (nine CDMs total) that progressively produce $1024.0\times1024.0$, $6400.0\times6400.0$, and $41344.0\times41344.0$ pixel WSIs, with stage-wise conditioning and inpainting. Evaluations on GLIOMA, BREAST, and KIDNEY show URCDMs outperforming state-of-the-art baselines on FID-10k, pFID-50k, and IP/IR, and expert blinding studies indicate generated WSIs are largely indistinguishable from real ones, albeit with some artefact sensitivity in certain cases. The approach promises privacy-preserving, gigapixel histopathology data across scales with potential impact on education and downstream modeling, while future work will target efficiency, a shared multi-magnification CDM, and multi-modal extensions.

Abstract

Diagnosing medical conditions from histopathology data requires a thorough analysis across the various resolutions of Whole Slide Images (WSI). However, existing generative methods fail to consistently represent the hierarchical structure of WSIs due to a focus on high-fidelity patches. To tackle this, we propose Ultra-Resolution Cascaded Diffusion Models (URCDMs) which are capable of synthesising entire histopathology images at high resolutions whilst authentically capturing the details of both the underlying anatomy and pathology at all magnification levels. We evaluate our method on three separate datasets, consisting of brain, breast and kidney tissue, and surpass existing state-of-the-art multi-resolution models. Furthermore, an expert evaluation study was conducted, demonstrating that URCDMs consistently generate outputs across various resolutions that trained evaluators cannot distinguish from real images. All code and additional examples can be found on GitHub.

URCDM: Ultra-Resolution Image Synthesis in Histopathology

TL;DR

The paper tackles the challenge of generating coherent, ultra-high-resolution histopathology WSIs that preserve hierarchical structure across magnifications, addressing memory and privacy issues of patch-based diffusion. It introduces Ultra-Resolution Cascaded Diffusion Models (URCDMs), a three-stage cascade of diffusion processes (nine CDMs total) that progressively produce , , and pixel WSIs, with stage-wise conditioning and inpainting. Evaluations on GLIOMA, BREAST, and KIDNEY show URCDMs outperforming state-of-the-art baselines on FID-10k, pFID-50k, and IP/IR, and expert blinding studies indicate generated WSIs are largely indistinguishable from real ones, albeit with some artefact sensitivity in certain cases. The approach promises privacy-preserving, gigapixel histopathology data across scales with potential impact on education and downstream modeling, while future work will target efficiency, a shared multi-magnification CDM, and multi-modal extensions.

Abstract

Diagnosing medical conditions from histopathology data requires a thorough analysis across the various resolutions of Whole Slide Images (WSI). However, existing generative methods fail to consistently represent the hierarchical structure of WSIs due to a focus on high-fidelity patches. To tackle this, we propose Ultra-Resolution Cascaded Diffusion Models (URCDMs) which are capable of synthesising entire histopathology images at high resolutions whilst authentically capturing the details of both the underlying anatomy and pathology at all magnification levels. We evaluate our method on three separate datasets, consisting of brain, breast and kidney tissue, and surpass existing state-of-the-art multi-resolution models. Furthermore, an expert evaluation study was conducted, demonstrating that URCDMs consistently generate outputs across various resolutions that trained evaluators cannot distinguish from real images. All code and additional examples can be found on GitHub.
Paper Structure (5 sections, 2 equations, 3 figures, 3 tables)

This paper contains 5 sections, 2 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Evaluation and Results
  4. Conclusion
  5. Appendix

Figures (3)

  • Figure 1: Detailed overview of the URCDM image generation process. The medium and high-magnification CDMs are sampled many times, and the patches generated are stitched together; one sample is shown as an example. A blue outline indicates the lower-magnification conditioning image, to teach the context for the new generation process. A green outline indicates the resultant patch that will be 'zoomed in' on and generated by a baseline CDM. Red lines indicate the output of each magnified image. Not to scale.
  • Figure 2: Random URCDM samples vs. Outpainting and StyleGAN. More qualitative examples can be found in the Appendix.
  • Figure 3: Samples of an ultra-resolution WSIs generated using a URCDM. Full-scale ultra-resolution image ($41344.0 \times 41344.0$ pixels), and highly zoomed-in crops of that image ($1024 \times 1024$ pixels) compared against random real and baseline images.