Room: AAPM ePoster Library
In current clinical practice, pretreatment dosimetric verification is an important step to ensure a planned dose being delivered. Although various measuring tools and verification methodologies have developed, 3D dose verification is still challenged. In this work, we report on a novel deep learning model to derive dose distributions at multiple depths from portal images collected from an in-house camera-based radioluminescence imaging system (CRIS).
Radioluminescent portal images collected from our CRIS and the corresponding dose distributions calculated in the TPS were acquired as the paired dataset in this study. A total of 696 image pairs with different beam shapes were used for network training. The proposed functional generative adversarial network (fGAN) was trained to convert portal image to corresponding plane dose at different depth-in-water. A series of regular beams and a clinic prostate IMRT beam were used for validation. The predicted dose maps were evaluated by comparing to the TPS calculations in terms of gamma pass rate and median gamma index with 2% / 2mm and 1% / 1mm acceptance criteria.
The averaged gamma pass rates were up to 100% and 98% for the regular field beams and a prostate IMRT beam, respectively, for depths of 1.5cm, 5cm and 10cm. The median gamma index is less than 0.4 for all the cases.
Leveraging from the proposed method: (1) dose distributions can be verified at multiple depths; (2) learning with unpaired images is supported to allow difference between the real dose and TPS calculations those were used as the training labels; (3) the glare artifacts in the camera-based radioluminescence imaging systems are alleviated.