No Cover Image

Conference Paper/Proceeding/Abstract 102 views 1 download

Guided Latent Diffusion for Universal Medical Image Segmentation

Mattia Salsi, Yunying Wang, Chen Hu, Hans Ren, Jingjing Deng, Xianghua Xie Orcid Logo

International Conference on AI-generated Content

Swansea University Authors: Chen Hu, Hans Ren, Xianghua Xie Orcid Logo

  • Diffusion_Segmentation.pdf

    PDF | Accepted Manuscript

    Author accepted manuscript document released under the terms of a Creative Commons CC-BY licence using the Swansea University Research Publications Policy (rights retention).

    Download (1.35MB)

Abstract

Deep learning based medical segmentation still presents a great challenge due to the lack of large-scale datasets in the medical domain. The existing publicly available datasets vary significantly in terms of imaging modalities and target anatomies. This paper presents a novel guided latent diffusio...

Full description

Published in: International Conference on AI-generated Content
Published:
URI: https://cronfa.swan.ac.uk/Record/cronfa68384
Abstract: Deep learning based medical segmentation still presents a great challenge due to the lack of large-scale datasets in the medical domain. The existing publicly available datasets vary significantly in terms of imaging modalities and target anatomies. This paper presents a novel guided latent diffusion model for universal medical segmentation, capable of segmenting diverse anatomical structures using a single and unified architecture. Given a Contrastive Language-Image Pretraining (CLIP) embedding specifying the target anatomical structure, the proposed model leverages a collection of datasets covering the diverse structures which can segment any anatomical targets that are presented in the aggregated data. By performing diffusion fully in latent space, we achieve comparable results to pixel-space diffusion with significantly lower computational cost. The proposed methods demonstrates competitive performance against existing deep learning-based discriminative approaches on several benchmarks. Furthermore, it shows strong generalization capabilities on unseen datasets.
College: Faculty of Science and Engineering

Similar Items