ACCELERATING DIFFUSION MODELS VIA PRESEGMENTATION DIFFUSION SAMPLING FOR MEDICAL IMAGE SEGMENTATION 论文阅读

Accelerating Diffusion Models via Pre-segmentation Diffusion Sampling for Medical Image Segmentation

Based on the Denoising Diffusion Probabilistic Model (DDPM), medical image segmentation can be described as a conditional image generation task, which allows to compute pixel-wise uncertainty maps of the segmentation and allows an implicit ensemble of segmentations to boost the segmentation performance.

http://arxiv.org/abs/2210.17408

Abstract

• Problem
DDPM requires many iterative denoising steps to generate segmentations from Gaussian noise, resulting in extremely inefficient inference.

• Method
The key idea is to obtain pre-segmentation results based on a separately trained segmentation network, and construct noise predictions (non-Gaussian distribution) according to the forward diffusion rule. We can then start with noisy predictions and use fewer reverse steps to generate segmentation results.

Method

Dataset