MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep Learning-Augmented Novel Radioluminescence Imaging System for Quality Assurance of Multileaf Collimators (MLC)

M jia*, X Li, Y Yang, L Wang, L Xing, Stanford University, Stanford, CA

Presentations

(Thursday, 7/16/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose:
Accurate multileaf collimator (MLC) positioning is vital to guarantee prescribed tumor dose conformity and minimize radiation to the organs at risk. The quality assurance (QA) of MLC is typically accomplished by analyzing portal images with sub-millimeter accuracy requirement. In this work, we report on a first volumetric camera-based radioluminescence imaging system (CRIS) with high image fidelity for reliable MLC QA.

Methods:
An in-house developed CRIS was used to acquire image data. In our CRIS, a scintillator sheet is overlaid on the inner surface a hollow cylinder. Leveraging from a hemispheric mirror, radiation beam projected to the cylindrical receptor could be captured for full gantry angles. A weakly supervised learning is proposed to synthesize benchmark EPID images from our CRIS measurement. The training labels were collected from a cine-mode EPID which supports high-speed acquisition while suffering from synchronization artifacts. A total of 696 image pairs with two basic beam shapes were used for network training. Our results are validated by comparing to benchmark EPID image, which is obtained by changing the EPID to integration mode with substantial dose delivered per irradiation. Image quality were evaluated in terms of PSNR, SSIM and gamma index.

Results:
Experimental results show that our system produces promising portal image comparable to an integration-mode EPID (up to 99% agreement with a 2% 2mm acceptance criteria) and achieves measurement sensitivity over submillimeter (0.3mm at least) MLC leaf displacement.

Conclusion:
In this work, (1) we propose the first volumetric camera-based radioluminescence imaging system (CRIS) for accurate MLC QA. (2) We propose a novel weakly supervised learning model for CRIS image enhancement to synthesize benchmark EPID images from raw CRIS measurement. The outstanding imaging performance was validated by conducting a range of experiments, highly inspiring the developed system as a potential low-cost alternative of general EPID.

Keywords

Radiation Detectors, Radiography, Scintillation Cameras

Taxonomy

IM- Optical : Development (new technology and techniques)

Contact Email