TransDose: a transformer-based UNet model for fast and accurate dose calculation for MR-LINACs

Abstract

Objective: To present a transformer-based UNet model (TransDose) for fast and accurate dose calculation for magnetic resonance-linear accelerators (MR-LINACs). Approach: A 2D fluence map from each beam was first projected into a 3D fluence volume and then fed into the TransDose model together with patient density volume and output predicted beam dose. The proposed TransDose model combined a 3D residual UNet with a transformer encoder, where convolutional layers extracted the volumetric spatial features, and the transformer encoder processed the long-range dependencies in a global space. Ninety-eight cases with four tumor sites (brain, nasopharynx, lung, and rectum) treated with fixed-beam intensity-modulated radiotherapy were included in the dataset; 78 cases were used for model training and validation; and 20 cases were used for testing. The ground-truth beam doses were calculated with Monte Carlo (MC) simulations within 1% statistical uncertainty and magnetic field strength B = 1.5 T in the superior and inferior direction. Beam angles from the training and validation datasets were rotated 2–5 times, and doses were recalculated to augment the datasets. Results: The dose-volume histograms and indices between the predicted and MC doses showed good consistency. The average 3D γ-passing rates (3%/2 mm, for dose regions above 10% of maximum dose) were 99.13±0.89% (brain), 98.31±1.92% (nasopharynx), 98.74±0.70% (lung), and 99.28±0.25% (rectum). The average dose calculation time, which included the fluence projection and model prediction, was less than 310 ms for each beam. Significance: We successfully developed a transformer-based UNet dose calculation model—TransDose in magnetic fields. Its accuracy and efficiency indicated its potential for use in online adaptive plan optimization for MR-LINACs.

Sub‐second photon dose prediction via transformer neural networks

“…In general, iDoTA achieves higher gamma pass rates than all previous convolutional models, also compared to the purely convolutional iDoTA-conv variant trained with identical dataset, training procedure and architecture (except for the transformer encoder). As in previous proton studies 39 and the concurrent TransDose, 40 these findings demonstrate that the addition of the transformer-being able to capture relationships between distant features, as opposed to convolutions-seems to be beneficial for dose prediction tasks. Moreover, our method outperforms the concurrent TransDose transformer model in both accuracy and speed.…”

“…All prediction times for all models include the time needed to generate and prepare the inputs, predict the output and (for full dose distributions) accumulate beam doses. For individual beam prediction, iDoTA is significantly faster than any other competitor, being 30–60x faster than the 3D U‐net models and 6x faster than the concurrent transformer model TransDose 40 . Likewise, iDoTA predicts full dose distribution from VMAT plans (with 194–354 beams per plan) on average in 8 s, representing a 10–80x speed‐up compared to the IMRT (with ≈ 10 beams) U‐net models.…”

Sub-second photon dose prediction via transformer neural networks

“…plans available, we compare the Γ(1 mm, 1%), Γ(2 mm, 2%) and Γ(3 mm, 3%) gamma pass rate to the values reported in previous studies. In particular, iDoTA's accuracy and inference times are compared to those of: convolutional U-net architectures predicting each beam in the plan individually [30,31]; convolutional models de-noising MC dose distributions [24,23]; and a concurrent 3D U-net and transformer model for MR-Linac dose prediction [39].…”

“…All prediction times for all models include the time needed to generate and prepare the inputs, predict the output and (for full dose distributions) accumulate beam doses. For individual beam prediction, iDoTA is significantly faster than any other competitor, being 30-60x faster than the 3D U-net models and 6x faster than the concurrent transformer model TransDose [39]. Likewise, iDoTA predicts full dose distribution from VMAT plans (with 194-354 beams per plan) on average in 8 seconds, representing a 10-80x speed-up compared to the IMRT (with ≈ 10 beams) U-net models.…”

“…In this study, we present a deep learning model that can predict dose distributions in few milliseconds with clinically acceptable accuracy. As in concurrent work [39], we harness the power of hybrid Transformer and 3D convolutional architectures, adapting the previous transformer-based proton dose calculation model [37] to predict the dose of much bigger photon broad beams. As shown in Figure 1, the proposed improved Dose Transformer Algorithm (iDoTA) combines a series of 3D convolutional layers modeling local dose and tissue variations, with a Transformer backbone routing information along the depth of the entire photon beam.…”