(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: IMRT QA measurements are routinely performed prior to treatment delivery to verify dose calculation and delivery accuracy. In this work, we use a novel deep convolution neural network (ConvNet) to predict IMRT patient QA passing rates for portal dosimetry.
Methods: 182 IMRT plans for various treatment sites were planned and delivered with portal dosimetry on two TrueBeam and two Trilogy LINACs. A total of 1497 beams were collected and analyzed per beam using gamma criteria of 2%/2mm with a 5% threshold. We train this ConvNet on 1259 beams with ten-fold cross-validation and test the model on the remaining 238 beams. Fluence maps calculated for each plan were used as inputs to a ConvNet. Instead of hand-designed ConvNet architectures, we used Neural Architecture Search (NAS) to search the best architecture for the ConvNet based on fluence maps. The performance of the ConvNet was compared with three tree-based regression models and two state-of-art hand-designed ConvNet models.
Results: The ConvNet architecture we found has a complex topological structure with 110 nodes and 132 edges. The model achieved 98.3% of predictions within 3% of the measured 2%/2mm gamma passing rates with a maximum error of 3.1% and a mean absolute error of 1.0%. Our results showed that this novel architecture search approach outperformed other machine-learning approaches with hand-crafted features and architecture.
Conclusion: We have developed an image-based ConvNet model using neural architecture search for IMRT QA prediction. We showed that QA passing rates can be accurately predicted based on the fluence maps and do not need the knowledge of manually extract relevant features and plan complexity metrics. The proposed approach will allow physicists to better identify the failures of IMRT QA measurements and to develop proactive QA approaches.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
