Related Experiment Video
Updated: Sep 3, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
A deep image-to-image network organ segmentation algorithm for radiation treatment planning: principles and
Sebastian Marschner1,2, Manasi Datar3, Aurélie Gaasch4
1Department of Radiation Oncology, University Hospital, LMU Munich, Munich, Germany. sebastian.marschner@med.uni-muenchen.de.
Radiation Oncology (London, England)
|July 22, 2022
Summary
A deep learning algorithm automates organ contouring for radiation therapy planning. This deep image-to-image network (DI2IN) shows high accuracy, simplifying and speeding up treatment planning for organs at risk.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Healthcare
- Radiation Oncology
Background:
- Radiation treatment planning requires accurate delineation of organs at risk (OARs) on computed tomography (CT) images.
- Manual contouring of OARs is time-consuming and subject to inter-observer variability.
Purpose of the Study:
- To develop and evaluate a deep network algorithm for automated OAR contouring in the thorax and pelvis.
- To assess the accuracy and efficiency of the algorithm compared to manual contouring.
Main Methods:
- A deep reinforcement learning technique identifies the region of interest (ROI) using anatomical landmarks.
- A deep image-to-image network (DI2IN) with a convolutional encoder-decoder architecture performs segmentation within the ROI.
- Evaluation involved comparing automated contours with manual contours (Radiation Therapy Oncology Group guidelines) on 237 thoracic and 102 pelvic CT datasets using Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD95).
Main Results:
- High correlations were observed between automated and manual contours.
- Excellent accuracy was achieved for lungs (DSC 0.97), heart (DSC 0.92), bladder (DSC 0.88), and rectum (DSC 0.79).
- Visual inspection confirmed excellent agreement, with minor exceptions for heart and rectum.
Conclusions:
- The DI2IN algorithm provides automated OAR contours comparable to expert manual delineations.
- This deep learning approach simplifies and accelerates the contouring process in radiation treatment planning.
- While generally accurate, some cases, particularly for heart and rectum, may still require manual adjustments.
