# Deep Learning-based Biological Anatomical Landmark Detection in Colonoscopy Videos

**Authors:** Kaiwei Che, Chengwei Ye, Yibing Yao, Nachuan Ma, Ruo Zhang, Jiankun Wang, Max Q. -H. Meng  
**arXiv:** [2108.02948](https://arxiv.org/abs/2108.02948)  
**Date:** 2021-08-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.02948v1/

## Key points

- Colonoscopy is a widely used imaging tool for examining the gastrointestinal tract and detecting lesion areas
- Automatic detection of biological anatomical landmarks within the colon is highly demanded to reduce clinician burden and improve patient outcomes
- A novel deep learning-based approach is proposed to detect these landmarks in colonoscopy videos
- The proposed approach involves pre-processing raw colonoscopy video sequences to reject interference frames and then utilizing a ResNet-101 based network to separately detect three biological anatomical landmarks and obtain intermediate detection results
- Post-processing of intermediate detection results is performed by identifying incorrectly predicted frames based on their temporal distribution and reassigning them back to the correct class
- The experimental results demonstrate that the proposed model achieves an average detection accuracy of 99.75% with an average Intersection over Union (IoU) of 0.91 between predicted landmark periods and ground truth, indicating a high degree of similarity between predicted and actual landmark periods in colonoscopy videos
- The findings suggest that the proposed model accurately detects and localizes biological anatomical landmarks from colonoscopy videos which could potentially reduce clinician burden and improve patient outcomes
- Authors include Kaiwei Che, Chengwei Ye, Yibing Yao, Nachuan Ma, Ruo Zhang, Jiankun Wang, and Max Q.-H. Meng

## Summary

Colonoscopy is a widely used imaging tool for examining the gastrointestinal (GI) tract and detecting lesion areas. To reduce the burden on clinicians and improve patient outcomes, automatic detection of biological anatomical landmarks within the colon is highly demanded. In this article, a novel deep learning-based approach is proposed to detect these landmarks in colonoscopy videos. The proposed approach involves pre-processing raw colonoscopy video sequences to reject interference frames and then utilizing a ResNet-101 based network to separately detect three biological anatomical landmarks and obtain intermediate detection results. Post-processing of intermediate detection results is performed by identifying incorrectly predicted frames based on their temporal distribution and reassigning them back to the correct class. The experimental results demonstrate that the proposed model achieves an average detection accuracy of 99.75% with an average Intersection over Union (IoU) of 0.91 between predicted landmark periods and ground truth. This indicates a high degree of similarity between predicted and actual landmark periods in colonoscopy videos. The findings suggest that the proposed model accurately detects and localizes biological anatomical landmarks from colonoscopy videos which could potentially reduce clinician burden and improve patient outcomes. The authors of this study include Kaiwei Che, Chengwei Ye, Yibing Yao, Nachuan Ma, Ruo Zhang, Jiankun Wang, and Max Q.-H. Meng whose work highlights the potential benefits of using deep learning-based approaches for detecting biological anatomical landmarks in medical imaging applications such as colonoscopy videos.