Data augmentation method for computer-aided diagnosis using specular reflection.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Colorectal cancer (CRC) screening relies on colonoscopy, but endoscopist skill variability impacts polyp detection and diagnosis.
• Advanced imaging like narrow-band imaging (NBI) aids real-time polyp assessment.
• Artificial intelligence (AI), specifically deep learning (DL), offers potential for computer-aided polyp detection and diagnosis (CADe/CADx).

Purpose of the Study:

• To develop and evaluate a novel data augmentation technique using specular reflection (SR) for enhancing DL models in colonoscopy.
• To improve the robustness and performance of DL-based computer-aided diagnosis (CADx) systems for colon polyps.
• To assess the effectiveness of SR augmentation compared to conventional methods, particularly under data-limited conditions.

Main Methods:

• Developed an SR augmentation method incorporating SR generation and inpainting.
• Integrated SR augmentation into conventional techniques.
• Trained and tested two DL architectures (CNN and Vision Transformer) on a dataset of 2,616 NBI colonoscopy images.
• Conducted stress tests using varying data usage ratios.

Main Results:

• SR augmentation demonstrated superior robustness and performance compared to other methods, especially with limited training data.
• DL models utilizing SR augmentation showed significantly improved accuracy in polyp diagnosis.
• The technique proved effective across different DL architectures (CNN and Vision Transformer).

Conclusions:

• Specular reflection data augmentation is a valuable technique for enhancing DL models in colonoscopy.
• This method improves the accuracy and reliability of AI-driven polyp diagnosis, particularly in data-scarce scenarios.
• SR augmentation supports the practical implementation of advanced CADx systems in clinical colonoscopy settings.