Related Experiment Video
Updated: Oct 21, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
DeepWL: Robust EPID based Winston-Lutz analysis using deep learning, synthetic image generation and optical
Michael John James Douglass1, James Alan Keal2
1School of Physical Sciences, University of Adelaide, Adelaide 5005, South Australia, Australia; Department of Medical Physics, Royal Adelaide Hospital, Adelaide 5000, South Australia, Australia.
Summary
A new deep learning model, DeepWL, analyzes Winston-Lutz (WL) quality assurance images from radiation therapy linear accelerators. Trained on synthetic data, DeepWL accurately localizes the radiation isocenter, potentially improving linac calibration accuracy.
Area of Science:
- Medical Physics
- Radiotherapy Technology
- Artificial Intelligence in Healthcare
Background:
- Accurate calibration of clinical linear accelerators is crucial for radiation therapy.
- The Winston-Lutz (WL) test is a key quality assurance (QA) test for localizing the radiation isocenter.
- Current methods for analyzing WL images can be time-consuming and require manual intervention.
Purpose of the Study:
- To develop and validate a novel deep learning model for analyzing EPID-based WL QA images.
- To introduce a new method for generating synthetic WL images and ground-truth masks using an optical path-tracing engine.
- To assess the performance of the deep learning model against existing methods for WL image analysis.
Main Methods:
- A deep learning model, DeepWL, was developed using Keras with a TensorFlow backend.
- Synthetic WL images and corresponding masks were generated using an optical path-tracing engine.
- DeepWL was trained on 1500 synthetic images with data augmentation for 180 epochs.
- The model's performance was evaluated on synthetic and real EPID-measured WL data.
Main Results:
- DeepWL achieved high segmentation accuracy on synthetic data, with mean Dice coefficients of 0.964 for ball bearings and 0.994 for MLC segments.
- When applied to EPID data, DeepWL's predicted mean displacements were statistically similar to the Canny Edge detection method.
- DeepWL showed better correlation with manual annotations for ball bearing localization compared to Canny Edge detection.
Conclusions:
- DeepWL demonstrates suitability for routine linac QA analysis.
- The deep learning approach offers potential advantages in segmentation robustness and displacement prediction accuracy over traditional methods.
- This work highlights the efficacy of using synthetic data and deep learning for medical imaging QA.