Denoising Diffusion Medical Models

TL;DR

This paper presents Denoising Diffusion Medical Models (DDMM), a generative approach that synthesizes realistic radiographical images and labels, improving biomedical image segmentation with limited annotated data.

Contribution

The introduction of DDMM, a novel diffusion-based model that generates paired medical images and labels, enhancing segmentation performance with minimal supervision.

Findings

DDMM produces realistic X-ray images and segmentations.

Using DDMM-augmented data improves segmentation accuracy.

The method outperforms other data-centric approaches.

Abstract

In this study, we introduce a generative model that can synthesize a large number of radiographical image/label pairs, and thus is asymptotically favorable to downstream activities such as segmentation in bio-medical image analysis. Denoising Diffusion Medical Model (DDMM), the proposed technique, can create realistic X-ray images and associated segmentations on a small number of annotated datasets as well as other massive unlabeled datasets with no supervision. Radiograph/segmentation pairs are generated jointly by the DDMM sampling process in probabilistic mode. As a result, a vanilla UNet that uses this data augmentation for segmentation task outperforms other similarly data-centric approaches.