Related Experiment Video
Updated: Dec 5, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
Improving CNN training on endoscopic image data by extracting additionally training data from endoscopic videos.
Georg Wimmer1, Michael Häfner2, Andreas Uhl1
1University of Salzburg, Department of Computer Sciences, Jakob-Haringerstrasse 2, Salzburg 5020, Austria.
Summary
This study introduces a novel method to increase labeled training data for convolutional neural networks (CNNs) in endoscopic image diagnosis by extracting real video data, significantly improving diagnostic performance.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Convolutional Neural Networks (CNNs) are crucial for computer-assisted endoscopic image diagnosis.
- A major limitation is the insufficient amount of labeled training data.
- Existing data augmentation techniques artificially generate data.
Purpose of the Study:
- To develop a technique for acquiring additional, real, labeled training data from endoscopic videos.
- To overcome the data scarcity problem in CNN-based endoscopic image analysis.
- To improve the performance of CNNs in diagnosing conditions like colonic polyps.
Main Methods:
- Extracting image patches from endoscopic videos based on initial labeled patches.
- Tracking polyp areas across video frames to identify new viewpoints and image qualities.
- Filtering extracted images for quality to ensure reliable training data.
- Utilizing real video data instead of artificial augmentation.
Main Results:
- Increased labeled image data by a factor of 39.
- Demonstrated clear and continuous performance improvement of CNNs using the proposed method.
- Acquired real image data under varied recording conditions (viewpoints, qualities).
Conclusions:
- The proposed method effectively addresses the limited training data issue in endoscopic diagnosis.
- Using real, varied data from videos enhances CNN performance significantly.
- This technique offers a practical solution for improving AI in medical imaging.
