BerDiff: Conditional Bernoulli Diffusion Model for Medical Image Segmentation 论文阅读

date
May 12, 2023
Last edited time
May 22, 2023 12:14 PM
status
Published
slug
BerDiff论文阅读
tags
DL
CV
summary
type
Post
Field
Plat

Abstract

这篇文章提出了一种新的医学图像分割方法,使用条件伯努利扩散模型。该模型可以实现准确和多样化的分割掩模,为临床医生提供有价值的参考。
我们首先提出使用伯努利噪声作为扩散核,而不是使用高斯噪声,以增强扩散模型对二元分割任务的能力,从而产生更准确的分割掩码。
notion image

Methodology

notion image
💡
eq4:
eq6:
优化的是:

Bernoulli forward process

在之前的扩散模型中,高斯噪声随着时间步长 的增加而逐渐增加。然而,对于分割任务,groundtruth masks 由离散值表示。为了解决这个问题,我们的 BerDiff 添加伯努利噪声。
where denotes the Bernoulli distribution with the probability parameters . Using the notation and , 我们可以在任意时间步 有效地对 进行采样,如下所示:
我们使用采样的伯努利噪声 重新参数化方程的 为 , ,其中 ⊕ 表示“异”的逻辑运算.
💡
这一步不太对, 前面的 哪去了

Diverse reverse process

本文的 Diverse reverse process 可以被视为一个马尔可夫链,从 Bernoulli 噪声 开始,并通过受输入医学图像 x 约束的中间潜变量来生成多样化的分割掩模。具体而言,BerDiff 的 Diverse reverse process 可以表示为:
其中,
被实现为:
表示绝对值运算。
💡
即使用 预测伯努利噪声.

Experiments

notion image
💡
Gaussian 扩散需要更长时间
notion image

Supp

notion image
notion image
notion image

© Lazurite 2021 - 2024