Weakly Supervised Polyp Segmentation in Colonoscopy Images Using Deep Neural Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colorectal cancer (CRC) is a major global health concern.
• Colonoscopy is crucial for CRC screening, but quality varies due to endoscopist skill and polyp characteristics.
• Deep learning shows promise for polyp identification in colonoscopy videos.

Purpose of the Study:

• To investigate the application of deep learning for real-time polyp segmentation in colonoscopy.
• To evaluate strategies for training segmentation models using weakly labeled data, specifically bounding boxes.
• To introduce and validate the Polyp-Box-Seg dataset for polyp segmentation research.

Main Methods:

• Development of a novel dataset (Polyp-Box-Seg) with 4070 colonoscopy images, including 1300 with precise segmentation masks.
• Training and evaluation of deep learning models for polyp segmentation using both precise masks and bounding box annotations.
• Implementation of a weakly supervised strategy leveraging bounding box data for model training.

Main Results:

• A model trained with segmentation masks achieved an 81.52% Dice coefficient.
• Weakly supervised training using bounding box annotations improved the Dice coefficient to 85.53%.
• The Polyp-Box-Seg dataset and a real-time segmentation system demonstration are publicly released.

Conclusions:

• Deep learning models can effectively perform real-time polyp segmentation in colonoscopy.
• Weakly supervised learning using bounding boxes is a viable and effective strategy for training polyp segmentation models.
• The developed dataset and system contribute to improving colonoscopy quality and CRC detection.