Related Experiment Video
Updated: Aug 5, 2026
Video Imaging and Spatiotemporal Maps to Analyze Gastrointestinal Motility in Mice
Published on: February 3, 2016
Causality-inspired representation learning with spatiotemporal memory for polyp detection in endoscopic videos
Zhuo Hu1, Changjin Sun2, Qi Zheng2
1School of Computer Science and Engineering, Southeast University, Nanjing, 210096, China; Laboratory of Image Science and Technology, School of Computer Science and Engineering, Nanjing, China; Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications, Nanjing, China.
This study introduces a novel framework for detecting colorectal polyps in colonoscopy videos. By using causality-inspired methods, it improves accuracy despite view shifts, enhancing early cancer detection.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Gastroenterology
Background:
- Early detection of colorectal polyps is vital for reducing colorectal cancer (CRC) morbidity and mortality.
- Colonoscopy video analysis is challenged by view shifts and environmental complexity, degrading visual cues and causing spurious correlations.
- Existing polyp detection methods struggle with the dynamic nature of colonoscopy footage.
Purpose of the Study:
- To develop a causality-inspired representation learning framework with spatiotemporal memory for robust polyp detection in colonoscopy videos.
- To address the challenge of view shifts and non-causal factors in polyp assessment during endoscopy.
- To improve the reliability and accuracy of automated polyp detection systems.
Main Methods:
- Proposed a causality-inspired representation learning framework (CIRL-Polyp) incorporating spatiotemporal memory.
- Introduced a View-Shift-Aware Causal Intervention Module (VACIM) to enforce prediction invariance under view shift perturbations.
- Developed a dual-branch detection framework with Causal Temporal Consistency Memory (CTCM) for enhanced temporal consistency and causal representation stabilization.
Main Results:
- CIRL-Polyp demonstrated superior performance compared to existing methods on public and private colonoscopy video datasets.
- The View-Shift-Aware Causal Intervention Module effectively removed non-causal influences caused by view shifts.
- The dual-branch framework and CTCM enhanced robustness by enforcing prediction consistency and leveraging long-range temporal dependencies.
Conclusions:
- The proposed CIRL-Polyp framework significantly improves polyp detection accuracy in colonoscopy videos by addressing view shift challenges from a causal perspective.
- The causality-inspired approach offers a more reliable method for automated polyp detection, potentially enhancing early diagnosis of colorectal cancer.
- The framework shows considerable clinical potential for improving colonoscopy-based screening and diagnosis.