S-CUDA: Self-cleansing unsupervised domain adaptation for medical image segmentation

Area of Science:

• Medical image analysis
• Deep learning in medical imaging
• Computer vision

Background:

• Deep convolutional neural networks (DCNNs) excel in medical image segmentation but are hindered by noisy labels and domain shift.
• Noisy labels arise from laborious and error-prone data annotation.
• Domain shift occurs due to data variations across different collection sites.

Purpose of the Study:

• To address the novel problem of unsupervised domain adaptation with noisy labeled data in medical image segmentation.
• To propose a robust algorithm, Self-Cleansing Unsupervised Domain Adaptation (S-CDUA), for simultaneous noise cleansing and domain adaptation.
• To improve the performance and generalization of deep learning models in realistic medical imaging scenarios.

Main Methods:

• Developed S-CDUA, a framework combining noisy-label learning and domain adaptation techniques.
• Employed two peer adversarial networks to identify and exchange high-confidence clean data, mitigating error accumulation and domain gap.
• Implemented a strategy for detecting and cleansing high-confidence noisy data for effective data recycling.

Main Results:

• S-CDUA demonstrated efficient noisy label cleansing and enhanced model generalization.
• The method achieved superior performance compared to state-of-the-art methods on optic disc/cup and spinal cord gray matter segmentation tasks.
• Experiments were conducted on REFUGE, Drishti-GS, and a multi-vendor spinal cord dataset.

Conclusions:

• The proposed S-CDUA method effectively handles both noisy labels and domain shift in medical image segmentation.
• This approach enables robust adaptation and prediction on target domains using real-world, imperfect data.
• S-CDUA offers a significant advancement for deep learning applications in medical imaging, improving segmentation accuracy and reliability.