Deep learning-augmented radiotherapy visualization with a cylindrical radioluminescence system.

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Image Processing

Background:

• External beam radiotherapy requires precise radiation delivery verification.
• On-board electronic portal imaging devices (EPIDs) have limitations in image quality due to glare and scattering.
• Existing methods for quality assurance can be costly and complex.

Purpose of the Study:

• To develop and validate a low-cost camera-based radioluminescence imaging system (CRIS) for high-quality beam visualization.
• To improve pre-treatment verification accuracy in external beam radiotherapy.
• To assess the system's capability for multi-leaf collimator (MLC) leaf-end position verification.

Main Methods:

• A deep learning model was trained using data from an on-board EPID to enhance image quality, mitigating artifacts like mirror glare and edge blurring.
• A cylindrical receptor was incorporated for independent measurement and co-planar detection.
• Three task-aware modules were integrated into the network for robustness during cine mode acquisition.
• The system was trained with various beam fields modulated by an MLC and validated on regular fields and IMRT cases.

Main Results:

• CRIS images demonstrated high similarity to high-quality EPID images, achieving a mean 2%/2 mm gamma pass rate of 99.14% for regular fields and 97.1% for IMRT cases.
• MLC leaf-end position verification using CRIS agreed with EPID measurements within 0.100 mm ± 0.072 mm.
• The system showed robust performance against artifacts common in EPID cine mode operation.

Conclusions:

• The developed low-cost camera-based radioluminescence imaging system (CRIS) provides accurate beam visualization and verification for radiotherapy.
• The deep learning-enhanced system effectively addresses image quality limitations of traditional EPIDs.
• CRIS offers a viable, cost-effective solution for routine quality assurance in radiation oncology.