Consolidated domain adaptive detection and localization framework for cross-device colonoscopic images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks for automatic polyp detection face performance degradation due to domain shifts from different scanners or hospitals.
• Domain adaptation is crucial for robust deployment of AI in colonoscopy for colorectal cancer screening.

Purpose of the Study:

• To propose a consolidated domain adaptive framework for effective detection and localization of polyps across diverse colonoscopic datasets.
• To address the domain gap in polyp data arising from variations in imaging devices and protocols.

Main Methods:

• A novel framework combining pixel-level adaptation (Gaussian Fourier Domain Adaptation - GFDA) and hierarchical feature-level adaptation (Hierarchical Attentive Adaptation - HAA, Iconic Concentrative Adaptation - ICA).
• GFDA unifies image styles via low-level spectrum replacement, reducing appearance discrepancies without content distortion.
• HAA and ICA minimize semantic and instance-level domain discrepancies, regularized by a Generalized Consistency Regularizer (GCR). A Centre Besiegement (CB) loss is used for localization.

Main Results:

• The proposed framework significantly outperforms existing domain adaptation detectors in polyp detection.
• Achieved a state-of-the-art recall rate of 87.5% for the polyp localization task.
• Demonstrated effective bridging of domain gaps between different colonoscopic datasets.

Conclusions:

• The consolidated domain adaptive framework enhances the robustness and accuracy of AI-based polyp detection and localization.
• This approach holds significant potential for improving colorectal cancer diagnosis and reducing mortality through more reliable AI tools.