A Source-Free Domain Adaptive Polyp Detection Framework With Style Diversification Flow.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automatic polyp detection is vital for colorectal cancer early diagnosis.
• Current deep learning methods often need extensive multi-site data or labeled source data for domain adaptation.
• Data privacy and storage limitations restrict the use of traditional unsupervised domain adaptation (UDA) techniques.

Purpose of the Study:

• To develop a source-free domain adaptive polyp detection method.
• To enable accurate polyp detection using only a pretrained model and unlabeled target data.
• To overcome limitations of existing UDA methods in data-scarce or privacy-sensitive scenarios.

Main Methods:

• Proposed Source Model as Proxy Teacher (SMPT) framework for source-free domain adaptation.
• Utilized Source Knowledge Distillation (SKD) and Proxy Teacher Rectification (PTR) for knowledge transfer and model refinement.
• Introduced Uncertainty-Guided Online Bootstrapping (UGOB) to handle domain-specific uncertainties.
• Developed Source Style Diversification Flow (SSDF) to enhance robustness against style variations.
• Integrated these components into an iterative optimization framework called SMPT++.

Main Results:

• Achieved state-of-the-art performance in cross-domain polyp detection across five datasets.
• Demonstrated superior performance compared to UDA methods that require access to source data.
• Validated the effectiveness of SMPT++ in improving polyp detection accuracy under challenging cross-domain conditions.

Conclusions:

• SMPT++ offers a powerful and practical solution for source-free domain adaptive polyp detection.
• The method effectively addresses data accessibility and privacy concerns in medical imaging.
• This advancement holds significant potential for improving colorectal cancer screening and diagnosis.