TSdetector: Temporal-Spatial self-correction collaborative learning for colonoscopy video detection.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Network (CNN)-based object detection models are increasingly used for polyp detection.
• Accurate polyp localization in colonoscopy videos remains challenging due to intra-sequence distribution heterogeneity and precision-confidence discrepancies.

Purpose of the Study:

• To propose a novel Temporal-Spatial self-correction detector (TSdetector) to address limitations in current polyp detection methods.
• To improve the continuous and accurate detection of polyps in colonoscopy videos.

Main Methods:

• Integration of temporal-level consistency learning and spatial-level reliability learning.
• Development of a global temporal-aware convolution to focus on global features between sequences.
• Implementation of a hierarchical queue integration mechanism for multi-temporal feature combination.
• Advancement of position-aware clustering for adaptive recalibration of prediction confidence and elimination of redundant bounding boxes.

Main Results:

• TSdetector achieves the highest polyp detection rate on three public colonoscopy video datasets.
• The proposed method outperforms existing state-of-the-art polyp detection techniques.
• Experimental validation demonstrates the effectiveness of the temporal and spatial learning components.

Conclusions:

• TSdetector offers a significant advancement in colonoscopy polyp detection by effectively handling sequence and confidence issues.
• The novel approach provides a more accurate and reliable tool for identifying polyps during endoscopic procedures.
• The developed model demonstrates superior performance compared to current methods, paving the way for enhanced diagnostic capabilities.