Spatio-temporal classification for polyp diagnosis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonoscopy is crucial for colorectal cancer (CRC) screening and polyp removal.
• Deep learning (DL) shows promise for computer-aided polyp characterization, aiding clinical decisions.
• Variability in polyp appearance during colonoscopy can reduce the stability of automated polyp classification.

Purpose of the Study:

• To investigate the utility of spatio-temporal information for enhancing polyp classification accuracy.
• To improve the robustness of deep learning models for distinguishing adenomatous from non-adenomatous colorectal lesions.
• To develop more reliable clinical decision support tools for colonoscopy procedures.

Main Methods:

• Implementation of two distinct methods utilizing spatio-temporal data for lesion classification.
• Extensive experimental validation on both internal and publicly available benchmark colonoscopy datasets.
• Focus on improving classification performance and model stability in the presence of visual variations.

Main Results:

• Demonstrated significant increases in classification performance through the integration of spatio-temporal features.
• Achieved enhanced robustness in polyp characterization, addressing the challenge of appearance variability.
• Validation across diverse datasets confirmed the effectiveness of the proposed approaches.

Conclusions:

• Spatio-temporal information integration significantly boosts the performance and reliability of DL-based polyp classification.
• The developed methods offer improved clinical decision support for adenoma detection during colonoscopy.
• This research contributes to more accurate and stable automated polyp analysis in CRC screening.