Alaka B G¹*, D Samuel², E Bentefour³, (1) Central University of Karnataka, Kalaburgi, Karnataka, IN, (2) Central University Of Karnataka, ,,(3) University of Maryland School of Medicine, Baltimore.

PO-GePV-M-17 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: To predict the WEPLs from energy-resolved dose functions (ERDF) using a deep learning algorithm.

Methods: A solid water wedge of dimension 30 cm x 10 cm x 30 cm in the path of the scanned proton beams is simulated using Geant4. The beam energy is varied between 100 MeV to 225 MeV in steps of 3 MeV and the exit dose is estimated from the simulation. This setup is used to obtain the calibration datasets (i.e, ERDF). This database is used for training a deep-learning model. The optimized model is used to predict the WEPL for box and staircase configurations (three steps stairs).

Results: The preliminary results show that the predicted WEPLs are in good agreement with the actual WEPLs for certain WEPL ranges. From the staircase results, the average error in predicting WEPL for 120 mm is 4.5% and the average error for 170 mm and 200 mm is 1.5%. However, the predicted WEPL varies significantly from the actual values in certain ranges of WEPL, and the work to achieve better accuracy is in progress. The advantage of the deep learning technique over the conventional technique is that the time taken for the deep learning model to predict the WEPL is considerably small (~ 1 minute to generate a 200 x 600 WEPL map).

Conclusion: Deep learning techniques offer a new avenue of analysis in the field of proton imaging especially in reducing the computation time. The technique needs to be fine-tuned with respect to the model parameters and features in order to predict the WEPLs within clinically acceptable limits.

Funding Support, Disclosures, and Conflict of Interest: Karnataka Science and Technology Promotion Society, Government of Karnataka.

ePosters

Keywords

Radiography, Dose, Simulation

Taxonomy

Not Applicable / None Entered.

Contact Email