Table of contents
- foreword
- MedSegDiff
- Dynamic Conditional Coding
- FF-Parser
- experiment
- MedSegDiff-V2
- overview
- Anchor Condition with Gaussian Spatial Attention
- Semantic Condition with SS-Former
- experiment
- Summarize
- reference
This article was first published on GiantPandaCV, and may not be reproduced without permission! !
foreword
In a previous article (ICLR 2023), we have introduced an application of the diffusion model in medical image segmentation (https://mp.weixin.qq.com/s/7g3_4hHfKCAp2WQibkbzgA), it is recommended that the diffusion model not Students who understand are given priority to read, which covers some basic concepts. The previous article was to apply diffusion to self-supervised learning, and MedSegDiff is a supervised framework, which has now been updated to the V2 version. The V2 version is different from V1 by using Transformer and is suitable for multi-classification. MedSegDiff-V1 has been accepted at MIDL 2023.
MedSegDiff
MedSegDiff introduces dynamic conditional encoding based on the original DPM to enhance DPM's step-by-step attention ability on medical image segmentation. The eigenfrequency parser (FF-Parser) can remove high-frequency noise in a given mask corrupted during segmentation. DPM is a generative model consisting of two stages, the forward