Human-level comparable control volume mapping with a deep unsupervised-learning model for image-guided radiation
Xiaokun Liang1, Maxime Bassenne1, Dimitre H Hristov1
1Department of Radiation Oncology, Stanford University, Stanford, CA, 94305, USA.
Computers in Biology and Medicine
|December 23, 2021
Summary
This study introduces a deep unsupervised learning method for precise patient positioning in head and neck cancer radiotherapy using control volume mapping. The new approach significantly improves registration accuracy compared to standard methods.
Area of Science:
- Medical Physics
- Radiotherapy
- Machine Learning
Background:
- Accurate patient positioning is critical for effective radiotherapy.
- Daily CT (dCT) and planning CT (pCT) scans are used for patient setup.
- Image registration methods aim to align dCT with pCT for accurate positioning.
Purpose of the Study:
- To develop a deep unsupervised learning method for precise patient positioning.
- To utilize control volume (CV) mapping from dCT to pCT.
- To automatically generate couch shifts (translation and rotation) for head and neck cancer (HNC) patients.
Main Methods:
- Proposed an unsupervised learning framework mapping CVs from dCT to pCT.
- Network inputs: dCT, pCT, and CV positions in pCT.
- Trained network to maximize image similarity between CVs in dCT and pCT.
- Evaluated on 554 CT scans from 158 HNC patients.
Main Results:
- System positioning errors: translation < 0.47 mm, rotation < 0.17°.
- Random positioning errors: translation < 1.13 mm, rotation < 0.29°.
- Improved registration within tolerance (2.0 mm/1.0°) from 66.67% to 90.91%.
Conclusions:
- Developed a deep unsupervised learning architecture for patient positioning using CV mapping.
- Method mitigates image artifact influence by differential weighting of CV regions.
- Achieved efficient and effective HNC patient positioning.
Keywords:
Head and neckImage registrationImage-guided radiation therapyPatient positioningUnsupervised learning
