Related Experiment Video
Updated: Jun 13, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Deep learning-based automatic segmentation of rectal tumors in endoscopic images.
Alana Thibodeau-Antonacci1, Aurélie Garant2, Corey Miller3
1Medical Physics Unit, Department of Oncology, McGill University, Montréal, Québec, Canada.
Brachytherapy
|June 11, 2026
Summary
Expert annotations for rectal tumors in endoscopic images show significant variability. A deep learning model improved segmentation but requires further refinement for clinical application in high-dose-rate brachytherapy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Endoscopic identification of rectal tumors is crucial but susceptible to observer error.
- High-dose-rate (HDR) brachytherapy relies on accurate tumor delineation.
Purpose of the Study:
- To assess inter- and intra-observer variability in delineating rectal lesions during HDR brachytherapy.
- To develop a deep learning model for automatic tumor segmentation in endoscopic images.
Main Methods:
- Three experts annotated tumors, scarring, ulcers, and radiation proctitis in 801 endoscopic images.
- Inter- and intra-observer variability were quantified using Dice scores.
- Four DeepLabV3 models were trained, including one on majority-vote contours.
- Model performance was evaluated on 60 unseen images.
Main Results:
- Tumor segmentation showed higher agreement (Dice: 0.83) compared to ulcers and proctitis (Dice: 0.36-0.57).
- Intra-observer variability yielded Dice scores ranging from 0.68 to 0.87.
- The majority-vote model achieved an average Dice score of 0.77 but produced false positives.
Conclusions:
- Expert annotations for rectal tumor segmentation in endoscopic images exhibit considerable variability.
- Automated contouring shows potential for AI-assisted brachytherapy, but further development is needed to address false positives.