Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ordinal Diffusion Models for Color Fundus Images

Gustav Schmidt, Philipp Berens, Sarah Müller

TL;DR

An ordinal latent diffusion model is proposed for generating color fundus images that explicitly incorporates the ordered structure of DR severity into the generation process, and interpolation experiments showed that the model captured a continuous spectrum of disease progression learned from ordered, coarse class labels.

Abstract

It has been suggested that generative image models such as diffusion models can improve performance on clinically relevant tasks by offering deep learning models supplementary training data. However, most conditional diffusion models treat disease stages as independent classes, ignoring the continuous nature of disease progression. This mismatch is problematic in medical imaging because continuous pathological processes are typically only observed through coarse, discrete but ordered labels as in ophthalmology for diabetic retinopathy (DR). We propose an ordinal latent diffusion model for generating color fundus images that explicitly incorporates the ordered structure of DR severity into the generation process. Instead of categorical conditioning, we used a scalar disease representation, enabling a smooth transition between adjacent stages. We evaluated our approach using visual realism metrics and classification-based clinical consistency analysis on the EyePACS dataset. Compared to a standard conditional diffusion model, our model reduced the Fréchet inception distance for four of the five DR stages and increased the quadratic weighted $κ$ from 0.79 to 0.87. Furthermore, interpolation experiments showed that the model captured a continuous spectrum of disease progression learned from ordered, coarse class labels.

Ordinal Diffusion Models for Color Fundus Images

TL;DR

An ordinal latent diffusion model is proposed for generating color fundus images that explicitly incorporates the ordered structure of DR severity into the generation process, and interpolation experiments showed that the model captured a continuous spectrum of disease progression learned from ordered, coarse class labels.

Abstract

It has been suggested that generative image models such as diffusion models can improve performance on clinically relevant tasks by offering deep learning models supplementary training data. However, most conditional diffusion models treat disease stages as independent classes, ignoring the continuous nature of disease progression. This mismatch is problematic in medical imaging because continuous pathological processes are typically only observed through coarse, discrete but ordered labels as in ophthalmology for diabetic retinopathy (DR). We propose an ordinal latent diffusion model for generating color fundus images that explicitly incorporates the ordered structure of DR severity into the generation process. Instead of categorical conditioning, we used a scalar disease representation, enabling a smooth transition between adjacent stages. We evaluated our approach using visual realism metrics and classification-based clinical consistency analysis on the EyePACS dataset. Compared to a standard conditional diffusion model, our model reduced the Fréchet inception distance for four of the five DR stages and increased the quadratic weighted from 0.79 to 0.87. Furthermore, interpolation experiments showed that the model captured a continuous spectrum of disease progression learned from ordered, coarse class labels.
Paper Structure (18 sections, 5 equations, 4 figures, 1 table)

This paper contains 18 sections, 5 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Ordinal-Aware Latent Diffusion Model
  5. Latent Autoencoding and Diffusion.
  6. Conditioning the Diffusion Model on Disease Stage and CFP Structure.
  7. Ordinal Disease Stage Conditioning.
  8. Evaluation of Image Realism and Disease Consistency
  9. Dataset and Preprocessing
  10. Training Details
  11. Results
  12. Ordinal Embedding Improved Realism and Disease Consistency
  13. Ordinal Diffusion Model Generates Images with Realistic Morphological Features and Disease Lesions
  14. Structure Encoder Enables Counterfactual Image-to-Image Generation
  15. Ordinal Diffusion Model Captures Continuous Disease Structure
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Overview figure of the proposed approach. Our diffusion model generates fundus images with different disease severity (b) by conditioning on ordinal embeddings (a).
  • Figure 2: CFPs generated from the ordinal diffusion model with equidistant margins are realistic and capture disease related lesions.
  • Figure 3: The structure encoder enabled counterfactual image-to-image generation. Original CFP of healthy eye on the left, generated images with ascending disease severity to the right. Below: difference image for each generated image from the healthy image.
  • Figure 4: The ordinal diffusion model learns the continuous disease spectrum underlying ordinal labels. Images generated at intermediate class values for (a) the equidistant margin model and (b) the learned margins model were generated with a multi-class DR classifier. Resulting label proportions are shown.