Active Domain Adaptation Of Medical Images Using Feature Disentanglement

Dwarikanath Mahapatra, Ruwan Tennakoon

Primary Area: representation learning for computer vision, audio, language, and other modalities

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.

Keywords: Active domain adaptation, feature disentanglement, chest xray, histopathology

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

TL;DR: Method for active domain adaptation of medical images

Abstract: State-of-the-art deep learning models often fail to generalize in the presence of distribution shifts between training (source) data and test (target) data. Domain adaptation techniques have been developed to address this challenge, leveraging either labeled data (supervised domain adaptation) or unlabeled data (unsupervised domain adaptation). The careful selection of target domain samples can significantly enhance model performance and robustness, while also reducing the overall data requirements. Active learning, a strategy for intelligently choosing informative samples with minimal annotation effort, offers a means to maximize performance. In this paper, we introduce an innovative method for active learning in the presence of domain shifts. We propose a novel feature disentanglement approach to decompose image features into domain-specific and task-specific components. Thereafter we define multiple novel cost functions that identify informative samples under domain shift. We test our proposed method for medical image classification using one histopathology dataset and two chest x-ray datasets. Experiments show our proposed approach achieves state-of-the-art performance when compared to both domain adaptation methods and other active domain adaptation techniques.

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: 7450