Enhancing polyp classification: A comparative analysis of spatio-temporal techniques.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colorectal cancer (CRC) is a significant global health issue, with early detection of pre-cancerous adenomatous polyps being critical for prevention.
• Narrow-band imaging (NBI) colonoscopy enhances polyp visualization, and artificial intelligence (AI) can improve polyp characterization during endoscopy.

Purpose of the Study:

• To compare the performance of three deep learning architectures (time-distributed 2D CNN-LSTM, 3D CNN, and 3D CNN-ConvLSTM2D) for colon polyp classification.
• To evaluate the effectiveness of incorporating spatiotemporal features for enhanced polyp characterization using NBI colonoscopy videos.
• To assess the generalizability and robustness of the best-performing model through cross-dataset validation.

Main Methods:

• Utilized three deep learning models: time-distributed 2D CNN-LSTM, 3D CNN, and a hybrid 3D CNN-ConvLSTM2D.
• Trained and evaluated models on a real-world clinical dataset of NBI colonoscopy videos from 60 patients in India (64 polyps).
• Performed cross-dataset validation on a public dataset to confirm model generalizability.

Main Results:

• The 3D CNN-ConvLSTM2D model demonstrated superior performance across all evaluation metrics compared to the other two architectures.
• Achieved a mean Negative Predictive Value (NPV) of 92%, exceeding the PIVI guidelines threshold for reliable polyp diagnosis.
• Showcased significant improvements in NPV and overall performance, with reduced false positives, compared to existing methods.

Conclusions:

• Incorporating spatiotemporal features via deep learning, particularly the 3D CNN-ConvLSTM2D model, is effective for accurate colon polyp classification.
• The developed model shows strong potential for real-world clinical application in improving polyp diagnosis and aiding CRC prevention.
• This study is the first to exclusively use an NBI polyp dataset to investigate the effectiveness of spatiotemporal information for polyp classification.