Author: Ricardo Garcia Santiago, Narges Miri, Daryl P. Nazareth, Ankit Pant, Mukund Seshadri 👨🔬
Affiliation: Roswell Park Comprehensive Cancer Center 🌍
Purpose: To develop a transformer-based deep learning network framework for predicting VMAT dose distributions. This can provide fast and efficient calculations with accuracies potentially comparable to those of state-of-the-art dose calculation methods.
Methods: A transformer network architecture was trained on the mapping, y=f(x,r), where x represents the CT image, r represents the aperture shape volume and y is the corresponding output dose distribution (the model’s prediction) using a modified version of the Improved Dose Transformer Algorithm (iDoTA). Training was performed with 800 VMAT control point datasets (two degree sub-arcs) obtained from four 6 MV clinical brain cases, with training run for four cycles with 120 epochs. For each control point, the aperture shape calculation was performed by averaging the raytracing of two aperture shapes to the patient volume, along with first-principles dose calculations. The patient plan CT was cropped/mapped to the dose matrix and HU values also converted to relative electron density. Ground truth (for training and testing) was determined using EGSnrc Monte Carlo platform. Evaluations were performed using 704 separate brain plan control point datasets.
Results: The deep learning framework evaluation showed that the input pairs could correlate to output dose with inference times of 15-30 ms per control point. The average gamma pass rate (3%, 3 mm, 10% global threshold) for the predicted control point dose matrices was 96.9%±1.3% and for full VMAT arc dose was 99.5%±0.1%. Visual colorwash/isodose displays indicate that relative dose distributions were accurately predicted by our method.
Conclusion: We have trained a deep learning framework based on transformer architecture to predict dose distributions from clinical VMAT plans with accuracy comparable to Monte Carlo/Acuros XB calculations. A complete arc (with 179) can be calculated within 12-14 seconds. Future work will include extending the training and evaluation sites to head and neck, lung, and prostate cases.