Deep reconstruction-recoding network for unsupervised domain adaptation and multi-center generalization in...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) excel at colonoscopy polyp detection but suffer performance degradation due to domain shift.
• Variations in imaging parameters, endoscope models, and lighting conditions (e.g., white light vs. narrow band imaging) create domain shift challenges.
• Retraining DNNs with new labeled data for each deployment is costly and time-consuming.

Purpose of the Study:

• To develop a domain adaptation model that enhances the cross-domain generalization of DNN-based polyp detection.
• To improve the adaptability of colonoscopy polyp detection models to diverse clinical settings without requiring additional labeled data.

Main Methods:

• Proposed the Deep Reconstruction-Recoding Network (DRRN), a domain adaptation model.
• DRRN jointly trains a supervised object detection network on labeled source data and an unsupervised reconstruction-recoding network on unlabeled target data.
• The model learns a shared encoding representation to reduce feature space distribution differences between domains.

Main Results:

• Evaluated DRRN on cross-domain datasets, demonstrating improved performance on the target domain compared to source-only training.
• Feature statistics and visualizations confirmed DRRN's ability to learn common distributions and feature invariance.
• The model effectively reduced the distribution differences between source and target feature spaces.

Conclusions:

• The DRRN significantly improves cross-domain polyp detection in colonoscopy.
• Generalization performance of DNN-based polyp detection models is enhanced without needing new labeled data.
• DRRN facilitates easier deployment of polyp detection models across different hospitals and endoscope types.