Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Ordinal Diffusion Model for Generating Medical Images with Different Severity Levels

Shumpei Takezaki, Seiichi Uchida

TL;DR

This work introduces an Ordinal Diffusion Model (ODM) to generate medical images with ordinal severity labels by enforcing monotonic relationships among class-conditioned noise estimates. ODM augments the standard diffusion objective with an ordinal loss and a time-weighted term, promoting interpolation/extrapolation across neighboring severity classes. Experimental results on EyePACS (DR severity) and LIMUC (UC Mayo score) show ODM achieving superior FID and fidelity metrics, especially in high-severity, data-scarce classes, indicating practical value for data augmentation and robust generation under ordinal constraints. The approach advances diffusion-based medical image synthesis by explicitly modeling ordinal structure, with potential impact on diagnostics support and severity-aware synthetic data generation.

Abstract

Diffusion models have recently been used for medical image generation because of their high image quality. In this study, we focus on generating medical images with ordinal classes, which have ordinal relationships, such as severity levels. We propose an Ordinal Diffusion Model (ODM) that controls the ordinal relationships of the estimated noise images among the classes. Our model was evaluated experimentally by generating retinal and endoscopic images of multiple severity classes. ODM achieved higher performance than conventional generative models by generating realistic images, especially in high-severity classes with fewer training samples.

An Ordinal Diffusion Model for Generating Medical Images with Different Severity Levels

TL;DR

This work introduces an Ordinal Diffusion Model (ODM) to generate medical images with ordinal severity labels by enforcing monotonic relationships among class-conditioned noise estimates. ODM augments the standard diffusion objective with an ordinal loss and a time-weighted term, promoting interpolation/extrapolation across neighboring severity classes. Experimental results on EyePACS (DR severity) and LIMUC (UC Mayo score) show ODM achieving superior FID and fidelity metrics, especially in high-severity, data-scarce classes, indicating practical value for data augmentation and robust generation under ordinal constraints. The approach advances diffusion-based medical image synthesis by explicitly modeling ordinal structure, with potential impact on diagnostics support and severity-aware synthetic data generation.

Abstract

Diffusion models have recently been used for medical image generation because of their high image quality. In this study, we focus on generating medical images with ordinal classes, which have ordinal relationships, such as severity levels. We propose an Ordinal Diffusion Model (ODM) that controls the ordinal relationships of the estimated noise images among the classes. Our model was evaluated experimentally by generating retinal and endoscopic images of multiple severity classes. ODM achieved higher performance than conventional generative models by generating realistic images, especially in high-severity classes with fewer training samples.
Paper Structure (17 sections, 8 equations, 4 figures, 1 table)

This paper contains 17 sections, 8 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Diagnostic Models with Ordinal Classes
  4. Diffusion Models for Medical Images
  5. Ordinal diffusion model
  6. Standard Diffusion Models ddpm
  7. Training the Model with Ordinal Relationship
  8. Experimental Results
  9. Experimental Setup
  10. Dataset
  11. Implimentation Details
  12. Evaluation Metrics and Comparative Methods
  13. Qualitative Evaluation of Generated Images
  14. Quantitative Evaluation of Generated Images
  15. Conculusion
  16. ...and 2 more sections

Figures (4)

  • Figure 1: (a) Endoscopic images with four ordinal severity classes of UC. (b) An overview of the proposed ordinal diffusion model. Its denoising process generates images while reflecting ordinal class relationships.
  • Figure 2: The effect of the ordinal relationship loss. Class indices $p$, $q$, and $r$ satisfy $1\leq p<q<r\leq C$. Our loss function forces the noise images to lie in a straight line while maintaining their ordinal relationship.
  • Figure 3: Examples of real images and generated images by ODM and the standard diffusion model (DM).
  • Figure 4: Class-wise FID by a standard diffusion model (DM) and ODM.