Meta-Learning for Bootstrapping Medical Image Segmentation from Imperfect Supervision

Qingyue Wei, Lequan Yu, Xianhang Li, Wei Shao, Cihang Xie, Lei Xing, Yuyin Zhou

Keywords: Semi-supervised learning, Meta-learning, Noisy labeling, Medical image segmentation

Abstract: Medical imaging has witnessed remarkable progress but usually requires a large amount of high-quality annotated data which is time-consuming and costly to obtain. To alleviate the annotation burden, learning from imperfect supervision (scarce or noisy annotations) has received much attention recently. In this paper, we present Meta-Learning for Bootstrapping Medical Image Segmentation (MLB-Seg), a unified meta-learning framework to sufficiently exploit the potential of imperfect supervision for medical image segmentation. In the face of noisy labeled data and unlabeled data, we first learn a segmentation model from a small clean set to generate initial labels for the unlabeled data and then gradually leverage the learner’s own predictions (i.e., the online pseudo labels) to bootstrap itself up via meta-learning. Specifically, MLB-Seg learns to dynamically assign per-pixel weight maps to both the imperfect labels (including both the generated labels and the noisy labels), as well as the pseudo labels commensurately to facilitate the bootstrapping procedure, where the weights are determined in a meta-process. To further improve the quality of the pseudo labels, we apply a consistency-based Pseudo Label Enhancement (PLE) scheme by ensembling predictions from various augmented versions of the same input. Noticing that the weight maps from these augmented variants can be extremely noisy from the meta-update, mean teacher is introduced into PLE to stabilize the weight map generation from the student (target) meta-learning model. Extensive experimental results on the public atrial and prostate segmentation datasets demonstrate that our method 1) achieves the state-of-the-art result under both semi- and noisy- supervision; 2) is robust against various imperfect supervisions. Code is publicly available at https://anonymous.4open.science/r/MLB-Seg-C80E.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning

TL;DR: A meta-based learning method for medical image segmentation under imperfect supervision