Few-shot Learning for Cardiac Segmentation via Self-supervised Multi-task LearningDownload PDF

06 Apr 2021 (modified: 16 May 2023)Submitted to MIDL 2021Readers: Everyone
Keywords: Self-supervised Learning, Multi-task Learning, Multi-modality CMR, Cardiac Segmentation
TL;DR: In this paper, we propose two novel self-supervised tasks, namely spatial encoding and modality encoding, and integrated the multiple SSL tasks to the frameworks of multi-task learning strategy.
Abstract: In this paper, we propose two novel self-supervised tasks, which encode the spatial and modality appearance information of cardiac images, respectively. Furthermore, we propose to ensemble multiple self-supervised tasks in a multi-task learning framework to learn more effective semantic representation for cardiac segmentation. The proposed approach was validated on the late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) images from the public MICCAI 2019 MSCMRseg Dataset, and was compared with two popular self-supervised tasks, including context inpainting (CI) and context restoration (CR). Results demonstrated that the proposed self-supervised tasks, as well as hybridized multi-task learning strategies, are effective in few-shot cardiac segmentation.
Paper Type: methodological development
Primary Subject Area: Segmentation
Secondary Subject Area: Unsupervised Learning and Representation Learning
Paper Status: original work, not submitted yet
Source Code Url: https://github.com/JiahangXu/spatial-encoding_and_modality-encoding
Data Set Url: http://www.sdspeople.fudan.edu.cn/zhuangxiahai/0/mscmrseg19/, https://acdc.creatis.insa-lyon.fr/description/index.html
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
3 Replies

Loading