DeepDose: Towards a fast dose calculation engine for radiation therapy using deep learning

Abstract

We present DeepDose, a deep learning framework for fast dose calculations in radiation therapy. Given a patient anatomy and linear-accelerator IMRT multi-leaf-collimator shape or segment, a novel set of physics-based inputs is calculated that encode the linac machine parameters into the underlying anatomy. These inputs are then used to train a deep convolutional network to derive the dose distribution of individual MLC shapes on a given patient anatomy. In this work we demonstrate the proof-of-concept application of DeepDose on 101 prostate patients treated in our clinic with fixed-beam IMRT. The ground-truth data used for training, validation and testing of the prediction were calculated with a state-of-the-art Monte Carlo dose engine at 1% statistical uncertainty per segment. A deep convolution network was trained using the data of 80 patients at the clinically used 3 mm3 grid spacing while 10 patients were used for validation. For another 11 independent test patients, the network was able to accurately estimate the segment doses from the clinical plans of each patient passing the clinical QA when compared with the Monte Carlo calculations, yielding on average 99.9%±0.3% for the forward calculated patient plans at 3%/3 mm gamma tests. Dose prediction using the trained network was very fast at approximately 0.9 seconds for the input generation and 0.6 seconds for single GPU inference per segment and 1 minute per patient in total. The overall performance of this dose calculation framework in terms of both accuracy and inference speed, makes it compelling for online adaptive workflows where fast segment dose calculations are needed.