Subtype-aware Unsupervised Domain Adaptation for Medical Diagnosis

Liu, Xiaofeng; Liu, Xiongchang; Hu, Bo; Ji, Wenxuan; Xing, Fangxu; Lu, Jun; You, Jane; Kuo, C.-C. Jay; Fakhri, Georges El; Woo, Jonghye

doi:10.48550/arxiv.2101.00318

Cited by 7 publications

(2 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…General UDA approaches [46,51,69,172] necessitate labeled samples from the source domain and unlabeled samples from the target domain for model training.…”

Section: Introductionmentioning

confidence: 99%

“…This discrepancy causes the performance of a network trained on one domain (source) to deteriorate when applied to data from another domain (target). Unsupervised Domain Adaptation (UDA) [46,51,69,172,190,191] seeks to minimize domain shift between source and target data, while avoiding the need for expensive pixel-level annotations. The source domain data is crucial in allowing the model to preserve valuable knowledge and iteratively reduce crossdomain differences during adaptation.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Domain adaptation in biomedical engineering: unsupervised, source-free, and black box approaches

Yuan

View full text Add to dashboard Cite

The remarkable advancements in deep learning methodologies over recent years can be attributed to the availability of large, high-quality labeled datasets, intricate network structures, and the swift progression of hardware technologies. However, in the biomedical engineering, the scarcity of data often poses a significant challenge in training a well-generalized model. This thesis delves into the potential of transfer learning methods as a viable solution to this data scarcity issue. The research provides a comprehensive exploration of Transfer Learning (TL), Unsupervised Domain Adaptation (UDA), Source-Free Domain Adaptation (SFDA), and Black-Box Domain Adaptation (BBDA) within the realm of biomedical engineering. It tackles the issue of data dependency in deep learning, with a particular emphasis on the application of TL in fine-tuning neural networks for specialized tasks. The study further investigates UDA techniques, specifically in the context of cross-dataset EEG classification tasks, and proposes entropy minimization-based methodologies to alleviate domain shifts. The thesis also scrutinizes SFDA in the context of medical imaging segmentation and abnormal ECG classification, introducing three innovative methodologies to overcome inherent challenges. Expanding on SFDA, the thesis evaluates the efficiency and accuracy of BBDA in real-time applications for cross-subject EEG driver drowsiness detection. This research contributes novel insights into handling data dependency, domain shifts, and privacy concerns in biomedical engineering, advocating for a nuanced approach where the scope of achievements is judiciously balanced against the challenges posed by scarce data.

show abstract

“…General UDA approaches [46,51,69,172] necessitate labeled samples from the source domain and unlabeled samples from the target domain for model training.…”

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Domain adaptation in biomedical engineering: unsupervised, source-free, and black box approaches

Yuan

View full text Add to dashboard Cite

show abstract

Mutual Information Regularized Identity-aware Facial ExpressionRecognition in Compressed Video

Liu¹,

Jin²,

Han³

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

This paper targets to explore the inter-subject variations eliminated facial expression representation in the compressed video domain. Most of the previous methods process the RGB images of a sequence, while the off-the-shelf and valuable expression-related muscle movement already embedded in the compression format. In the up to two orders of magnitude compressed domain, we can explicitly infer the expression from the residual frames and possible to extract identity factors from the I frame with a pre-trained face recognition network. By enforcing the marginal independent of them, the expression feature is expected to be purer for the expression and be robust to identity shifts. Specifically, we propose a novel collaborative min-min game for mutual information (MI) minimization in latent space. We do not need the identity label or multiple expression samples from the same person for identity elimination. Moreover, when the apex frame is annotated in the dataset, the complementary constraint can be further added to regularize the feature-level game. In testing, only the compressed residual frames are required to achieve expression prediction. Our solution can achieve comparable or better performance than the recent decoded image based methods on the typical FER benchmarks with about 3 times faster inference. tion, CAS (2017264), PolyU Central Research Grant G-YBJW, and Hong Kong Government General Research Fund GRF (Ref. No.152202/14E) are greatly appreciated.

show abstract

Adapting Off-the-Shelf Source Segmenter for Target Medical Image Segmentation

Liu¹,

Xing²,

Yang³

et al. 2021

Preprint

View full text Add to dashboard Cite

Unsupervised domain adaptation (UDA) aims to transfer knowledge learned from a labeled source domain to an unlabeled and unseen target domain, which is usually trained on data from both domains. Access to the source domain data at the adaptation stage, however, is often limited, due to data storage or privacy issues. To alleviate this, in this work, we target source free UDA for segmentation, and propose to adapt an "off-the-shelf" segmentation model pre-trained in the source domain to the target domain, with an adaptive batch-wise normalization statistics adaptation framework. Specifically, the domainspecific low-order batch statistics, i.e., mean and variance, are gradually adapted with an exponential momentum decay scheme, while the consistency of domain shareable high-order batch statistics, i.e., scaling and shifting parameters, is explicitly enforced by our optimization objective. The transferability of each channel is adaptively measured first from which to balance the contribution of each channel. Moreover, the proposed source free UDA framework is orthogonal to unsupervised learning methods, e.g., self-entropy minimization, which can thus be simply added on top of our framework. Extensive experiments on the BraTS 2018 database show that our source free UDA framework outperformed existing source-relaxed UDA methods for the cross-subtype UDA segmentation task and yielded comparable results for the cross-modality UDA segmentation task, compared with a supervised UDA methods with the source data.

show abstract

Subtype-aware Unsupervised Domain Adaptation for Medical Diagnosis

Cited by 7 publications

References 42 publications

Domain adaptation in biomedical engineering: unsupervised, source-free, and black box approaches

Domain adaptation in biomedical engineering: unsupervised, source-free, and black box approaches

Mutual Information Regularized Identity-aware Facial ExpressionRecognition in Compressed Video

Adapting Off-the-Shelf Source Segmenter for Target Medical Image Segmentation

Contact Info

Product

Resources

About