UniSeMi: Toward Unified Semi-supervised Medical Image Segmentation

Qingjie Zeng, Yutong Xie, Zilin LU, Mengkang Lu, Yong Xia

Keywords: medical image segmentation, semi-supervised learning

Abstract:

Semi-supervised learning (SSL) for medical image segmentation is put forward to mitigate the scarcity of annotation by leveraging unlabeled data. Recently proposed SSL works focus on designing task-specific models that process different tasks separately. This results in marginal improvement due to inadequate supervision from scarce labels of each single task.
To address this, we advocate learning a \textbf{Uni}fied \textbf{Se}mi-supervised segmentation model for \textbf{M}edical \textbf{i}maging (UniSeMi) by augmenting the label space, in which all pertinent task data are leveraged simultaneously. Specifically, UniSeMi can complete various missions using one single model with a task-prompted dynamic head. Beyond that, UniSeMi can learn from unlabeled data without requiring associated task information, $i.e.$, which task the unlabeled data belong to remains unknown.
To achieve this, we first synthesize an additional task by utilizing labeled data from pertinent tasks, and the synthetic task aims to instruct UniSeMi to be aware of all task semantics. In the context of unlabeled data learning, the aggregated prediction prompted by pertinent tasks is constrained to be consistent with the prediction prompted by the synthetic task, thus task information is not desired.
We evaluate UniSeMi on four public medical benchmarks, experiments show UniSeMi outperforms the second-best SSL method by 2.69\% and 8.92\% according to the averaged Dice and HD score, respectively. Code will be released.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

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.

Submission Number: 582