Sub-second photon dose prediction via transformer neural networks

DeepSMCP - Deep-learning powered denoising of Monte Carlo dose distributions within the Swiss Monte Carlo Plan

This work demonstrated the development of a fast, deep-learning framework (DeepSMCP) to mitigate noise in Monte Carlo dose distributions (MC-DDs) of photon treatment plans with high statistical uncertainty (SU) and its integration into the Swiss Monte Carlo Plan (SMCP). To this end, a two-channel input (MC-DD and computed tomography (CT) scan) 3D U-net was trained, validated and tested (80%/10%/10%) on high/low-SU MC-DD-pairs of 106 clinically-motivated VMAT arcs for 29 available CTs, augmented to 3074 pairs. The model was integrated into SMCP to enable a "one-click" workflow of calculating and denoising MC-DDs of high SU to obtain MC-DDs of low SU. The model accuracy was evaluated on the test set using Gamma passing rate (2% global, 2 mm, 10% threshold) comparing denoised and low-SU MC-DD. Calculation time for the whole workflow was recorded. Denoised MC-DDs match low-SU MC-DDs with average (standard deviation) Gamma passing rate of 82.9% (4.7%). Additional application of DeepSMCP to 12 unseen clinically-motivated cases of different treatment sites, including treatment sites not present during training, resulted in an average Gamma passing rate of 91.0%. Denoised DDs were obtained on average in 35.1 s, a 340-fold efficiency gain compared to low-SU MC-DD calculation. DeepSMCP presented a first seamlessly integrated promising denoising framework for MC-DDs.

A unified deep-learning framework for enhanced patient-specific quality assurance of intensity-modulated radiation therapy plans

BACKGROUND

Modern radiation therapy techniques, such as intensity-modulated radiation therapy (IMRT) and volumetric-modulated arc therapy (VMAT), use complex fluence modulation strategies to achieve optimal patient dose distribution. Ensuring their accuracy necessitates rigorous patient-specific quality assurance (PSQA), traditionally done through pretreatment measurements with detector arrays. While effective, these methods are labor-intensive and time-consuming. Independent calculation-based methods leveraging advanced dose algorithms provide a reduced workload but cannot account for machine performance during delivery.

PURPOSE

This study introduces a novel unified deep-learning (DL) framework to enhance PSQA. The framework can combine the strengths of measurement- and calculation-based approaches.

METHODS

A comprehensive artificial training dataset, comprising 400,000 samples, was generated based on a rigorous mathematical model that describes the physical processes of radiation transport and interaction within both the medium and detector. This artificial data was used to pretrain the DL models, which were subsequently fine-tuned with a measured dataset of 400 IMRT segments to capture the machine-specific characteristics. Additional measurements of five IMRT plans were used as the unseen test dataset. Within the unified framework, a forward prediction model uses plan parameters to predict the measured dose distributions, while the backward prediction model reconstructs these parameters from actual measurements. The former enables a detailed control point (CP)-wise analysis. At the same time, the latter facilitates the reconstruction of treatment plans from the measurements and, subsequently, dose recalculation in the treatment planning system (TPS), as well as an independent second check software (VERIQA). This method has been tested with an OD 1600 SRS and an OD 1500 detector array with distinct spatial resolution and detector arrangement in combination with a dedicated upsampling model for the latter.

RESULTS

The final models could deliver highly accurate predictions of the measurements in the forward direction and the actual delivered plan parameters in the backward direction. In the forward direction, the test plans reached median gamma passing rates better than 94% for the OD 1600 SRS measurements. The upsampled OD 1500 measurements show similar performance with similar median gamma passing rates but a slightly higher variability. The 3D gamma passing rates from the comparisons between the original and reconstructed dose distributions in patients lie between 95.4% and 98.2% for the OD 1600 SRS and 94.7% and 98.5% for the interpolated OD 1500 measurements. The dose volume histograms (DVH) of the original and the reconstructed plans, recalculated in both the TPS and VERIQA, were evaluated for the organs at risk and targets based on clinical protocols and showed no clinically relevant deviations.

CONCLUSIONS

The flexibility of the implemented model architecture allows its adaptability to other delivery techniques and measurement modalities. Its utilization also reduces the requirements of the measurement devices. The proposed unified framework could play a decisive role in automating QA workflow, especially in the context of real-time adaptive radiation therapy (ART).

Development and comprehensive clinical validation of a deep neural network for radiation dose modelling to enhance magnetic resonance imaging guided radiotherapy

Background and purpose

Online adaptive magnetic resonance imaging (MRI)-guided radiotherapy requires fast dose calculation algorithms to reduce intra-fraction motion uncertainties and improve workflow efficiency. While Monte-Carlo simulations are precise but computationally intensive, neural networks promise fast and accurate dose modelling in strong magnetic fields. This study aimed to train and evaluate a deep neural network for dose modelling in MRI-guided radiotherapy using a comprehensive clinical dataset.

Materials and methods

A dataset of 6595 clinical irradiation segments from 125 1.5 T MRI-Linac radiotherapy plans for various tumors sites was used. A 3D U-Net was trained with 3961 segments using 3D imaging data and field parameters as input, Root Mean Squared Error and a custom loss function, with full Monte-Carlo simulations as ground truth. For 2656 segments from 50 patients, gamma pass rates (γ-PR) for 3 mm/3%, 2 mm/2%, and 1 mm/1% criteria were calculated to assess dose modelling accuracy. Performance was also tested in a standardized water phantom to evaluate basic radiation physics properties.

Results

The neural network accurately modeled dose distributions in both patient and water phantom settings. Median (range) γ-PR of 97.7% (87.5-100.0%), 89.1% (69.7-99.4%), and 60.8% (38.5-82.1%) were observed for treatment plans, and 97.1% (55.5-100.0%), 88.8% (38.8-99.7%), and 61.7% (17.9-94.4%) for individual segments, across the three criteria.

Conclusion

High median γ-PR and accurate modelling in both water phantom and clinical data demonstrate the high potential of neural networks for dose modelling. However, instances of lower γ-PR highlight the need for comprehensive test data, improved robustness and future built-in uncertainty estimation.

Validation of a rapid algorithm for repeated intensity modulated radiation therapy dose calculations

As adaptive radiotherapy workflows and deep learning model training rise in popularity, the need for repeated applications of a rapid dose calculation algorithm increases. In this work we evaluate the feasibility of a simple algorithm that can calculate dose directly from MLC positions in near real-time. Given the necessary machine parameters, the intensity modulated radiation therapy (IMRT) doses are calculated and can be used in optimization, deep learning model training, or other cases where fast repeated segment dose calculations are needed. The algorithm uses normalized beamlets to modify a pre-calculated patient specific open field into any MLC segment shape. This algorithm was validated on 91 prostate IMRT plans as well as 20 lung IMRT plans generated for the Elekta Unity MR-Linac. IMRT plans calculated using the proposed method were found to match reference Monte Carlo calculated dose within98.02±0.84%and96.57±2.41%for prostate and lung patients respectively with a 3%/2 mm gamma criterion. After the patient-specific open field calculation, the algorithm can calculate the dose of a 9-field IMRT plan in 1.016 ± 0.284 s for a single patient or 0.264 ms per patient for a parallelized batch of 24 patients relevant for deep learning training. The presented algorithm demonstrates an alternative rapid IMRT dose calculator that does not rely on training a deep learning model while still being competitive in terms of speed and accuracy making it a compelling choice in cases where repetitive dose calculation is desired.

Quality assurance for online adaptive radiotherapy: a secondary dose verification model with geometry-encoded U-Net

In online adaptive radiotherapy (ART), quick computation-based secondary dose verification is crucial for ensuring the quality of ART plans while the patient is positioned on the treatment couch. However, traditional dose verification algorithms are generally time-consuming, reducing the efficiency of ART workflow. This study aims to develop an ultra-fast deep-learning (DL) based secondary dose verification algorithm to accurately estimate dose distributions using computed tomography (CT) and fluence maps (FMs). We integrated FMs into the CT image domain by explicitly resolving the geometry of treatment delivery. For each gantry angle, an FM was constructed based on the optimized multi-leaf collimator apertures and corresponding monitoring units. To effectively encode treatment beam configuration, the constructed FMs were back-projected to 30 cm away from the isocenter with respect to the exact geometry of the treatment machines. Then, a 3D U-Net was utilized to take the integrated CT and FM volume as input to estimate dose. Training and validation were performed on 381 prostate cancer cases, with an additional 40 testing cases for independent evaluation of model performance. The proposed model can estimate dose in ∼ 15 ms for each patient. The average γ passing rate ( 3 % / 2 mm , 10 % threshold) for the estimated dose was 99.9% ± 0.15% on testing patients. The mean dose differences for the planning target volume and organs at risk were 0.07 % ± 0.34 % and 0.48 % ± 0.72 % , respectively. We have developed a geometry-resolved DL framework for accurate dose estimation and demonstrated its potential in real-time online ART doses verification.

Application and progress of artificial intelligence in radiation therapy dose prediction

Radiation therapy (RT) nowadays is a main treatment modality of cancer. To ensure the therapeutic efficacy of patients, accurate dose distribution is often required, which is a time-consuming and labor-intensive process. In addition, due to the differences in knowledge and experience among participants and diverse institutions, the predicted dose are often inconsistent. In last several decades, artificial intelligence (AI) has been applied in various aspects of RT, several products have been implemented in clinical practice and confirmed superiority. In this paper, we will review the research of AI in dose prediction, focusing on the progress in deep learning (DL).

Deep learning based linear energy transfer calculation for proton therapy

Objective.This study aims to address the limitations of traditional methods for calculating linear energy transfer (LET), a critical component in assessing relative biological effectiveness (RBE). Currently, Monte Carlo (MC) simulation, the gold-standard for accuracy, is resource-intensive and slow for dose optimization, while the speedier analytical approximation has compromised accuracy. Our objective was to prototype a deep-learning-based model for calculating dose-averaged LET (LETd) using patient anatomy and dose-to-water (DW) data, facilitating real-time biological dose evaluation and LET optimization within proton treatment planning systems.Approach. 275 4-field prostate proton Stereotactic Body Radiotherapy plans were analyzed, rendering a total of 1100 fields. Those were randomly split into 880, 110, and 110 fields for training, validation, and testing. A 3D Cascaded UNet model, along with data processing and inference pipelines, was developed to generate patient-specific LETddistributions from CT images and DW. The accuracy of the LETdof the test dataset was evaluated against MC-generated ground truth through voxel-based mean absolute error (MAE) and gamma analysis.Main results.The proposed model accurately inferred LETddistributions for each proton field in the test dataset. A single-field LETdcalculation took around 100 ms with trained models running on a NVidia A100 GPU. The selected model yielded an average MAE of 0.94 ± 0.14 MeV cm-1and a gamma passing rate of 97.4% ± 1.3% when applied to the test dataset, with the largest discrepancy at the edge of fields where the dose gradient was the largest and counting statistics was the lowest.Significance.This study demonstrates that deep-learning-based models can efficiently calculate LETdwith high accuracy as a fast-forward approach. The model shows great potential to be utilized for optimizing the RBE of proton treatment plans. Future efforts will focus on enhancing the model's performance and evaluating its adaptability to different clinical scenarios.

Deep learning for high-resolution dose prediction in high dose rate brachytherapy for breast cancer treatment

Objective.Monte Carlo (MC) simulations are the benchmark for accurate radiotherapy dose calculations, notably in patient-specific high dose rate brachytherapy (HDR BT), in cases where considering tissue heterogeneities is critical. However, the lengthy computational time limits the practical application of MC simulations. Prior research used deep learning (DL) for dose prediction as an alternative to MC simulations. While accurate dose predictions akin to MC were attained, graphics processing unit limitations constrained these predictions to large voxels of 3 mm × 3 mm × 3 mm. This study aimed to enable dose predictions as accurate as MC simulations in 1 mm × 1 mm × 1 mm voxels within a clinically acceptable timeframe.Approach.Computed tomography scans of 98 breast cancer patients treated with Iridium-192-based HDR BT were used: 70 for training, 14 for validation, and 14 for testing. A new cropping strategy based on the distance to the seed was devised to reduce the volume size, enabling efficient training of 3D DL models using 1 mm × 1 mm × 1 mm dose grids. Additionally, novel DL architecture with layer-level fusion were proposed to predict MC simulated dose to medium-in-medium (Dm,m). These architectures fuse information from TG-43 dose to water-in-water (Dw,w) with patient tissue composition at the layer-level. Different inputs describing patient body composition were investigated.Main results.The proposed approach demonstrated state-of-the-art performance, on par with the MCDm,mmaps, but 300 times faster. The mean absolute percent error for dosimetric indices between the MC and DL-predicted complete treatment plans was 0.17% ± 0.15% for the planning target volumeV100, 0.30% ± 0.32% for the skinD2cc, 0.82% ± 0.79% for the lungD2cc, 0.34% ± 0.29% for the chest wallD2ccand 1.08% ± 0.98% for the heartD2cc.Significance.Unlike the time-consuming MC simulations, the proposed novel strategy efficiently converts TG-43Dw,wmaps into preciseDm,mmaps at high resolution, enabling clinical integration.

A deep learning based dynamic arc radiotherapy photon dose engine trained on Monte Carlo dose distributions

Background and purpose

Despite hardware acceleration, state-of-the-art Monte Carlo (MC) dose engines require considerable computation time to reduce stochastic noise. We developed a deep learning (DL) based dose engine reaching high accuracy at strongly reduced computation times.

Materials and methods

Radiotherapy treatment plans and computed tomography scans were collected for 350 treatments in a variety of tumor sites. Dose distributions were computed using a MC dose engine for ∼ 30,000 separate segments at 6 MV and 10 MV beam energies, both flattened and flattening filter free. For dynamic arcs these explicitly incorporated the leaf, jaw and gantry motions during dose delivery. A neural network was developed, combining two-dimensional convolution and recurrence using 64 hidden channels. Parameters were trained to minimize the mean squared log error loss between the MC computed dose and the model output. Full dose distributions were reconstructed for 100 additional treatment plans. Gamma analyses were performed to assess accuracy.

Results

DL dose evaluation was on average 82 times faster than MC computation at a 1 % accuracy setting. In voxels receiving at least 10 % of the maximum dose the overall global gamma pass rate using a 2 % and 2 mm criterion was 99.6 %, while mean local gamma values were accurate within 2 %. In the high dose region over 50 % of maximum the mean local gamma approached a 1 % accuracy.

Conclusions

A DL based dose engine was implemented, able to accurately reproduce MC computed dynamic arc radiotherapy dose distributions at high speed.

A Comparative Study of Deep Learning Dose Prediction Models for Cervical Cancer Volumetric Modulated Arc Therapy

Purpose: Deep learning (DL) is widely used in dose prediction for radiation oncology, multiple DL techniques comparison is often lacking in the literature. To compare the performance of 4 state-of-the-art DL models in predicting the voxel-level dose distribution for cervical cancer volumetric modulated arc therapy (VMAT). Methods and Materials: A total of 261 patients' plans for cervical cancer were retrieved in this retrospective study. A three-channel feature map, consisting of a planning target volume (PTV) mask, organs at risk (OARs) mask, and CT image was fed into the three-dimensional (3D) U-Net and its 3 variants models. The data set was randomly divided into 80% as training-validation and 20% as testing set, respectively. The model performance was evaluated on the 52 testing patients by comparing the generated dose distributions against the clinical approved ground truth (GT) using mean absolute error (MAE), dose map difference (GT-predicted), clinical dosimetric indices, and dice similarity coefficients (DSC). Results: The 3D U-Net and its 3 variants DL models exhibited promising performance with a maximum MAE within the PTV 0.83% ± 0.67% in the UNETR model. The maximum MAE among the OARs is the left femoral head, which reached 6.95% ± 6.55%. For the body, the maximum MAE was observed in UNETR, which is 1.19 ± 0.86%, and the minimum MAE was 0.94 ± 0.85% for 3D U-Net. The average error of the Dmean difference for different OARs is within 2.5 Gy. The average error of V40 difference for the bladder and rectum is about 5%. The mean DSC under different isodose volumes was above 90%. Conclusions: DL models can predict the voxel-level dose distribution accurately for cervical cancer VMAT treatment plans. All models demonstrated almost analogous performance for voxel-wise dose prediction maps. Considering all voxels within the body, 3D U-Net showed the best performance. The state-of-the-art DL models are of great significance for further clinical applications of cervical cancer VMAT.

The Quality Assurance of a 1.5 T MR-Linac

The recent introduction of a commercial 1.5 T MR-linac system has considerably improved the image quality of the patient acquired in the treatment unit as well as enabling online adaptive radiation therapy (oART) treatment strategies. Quality Assurance (QA) of this new technology requires new methodology that allows for the high field MR in a linac environment. The presence of the magnetic field requires special attention to the phantoms, detectors, and tools to perform QA. Due to the design of the system, the integrated megavoltage imager (MVI) is essential for radiation beam calibrations and QA. Additionally, the alignment between the MR image system and the radiation isocenter must be checked. The MR-linac system has vendor-supplied phantoms for calibration and QA tests. However, users have developed their own routine QA systems to independently check that the machine is performing as required, as to ensure we are able to deliver the intended dose with sufficient certainty. The aim of this work is therefore to review the MR-linac specific QA procedures reported in the literature.

A probabilistic deep learning model of inter-fraction anatomical variations in radiotherapy

Objective. In radiotherapy, the internal movement of organs between treatment sessions causes errors in the final radiation dose delivery. To assess the need for adaptation, motion models can be used to simulate dominant motion patterns and assess anatomical robustness before delivery. Traditionally, such models are based on principal component analysis (PCA) and are either patient-specific (requiring several scans per patient) or population-based, applying the same set of deformations to all patients. We present a hybrid approach which, based on population data, allows to predict patient-specific inter-fraction variations for an individual patient.Approach. We propose a deep learning probabilistic framework that generates deformation vector fields warping a patient's planning computed tomography (CT) into possible patient-specific anatomies. This daily anatomy model (DAM) uses few random variables capturing groups of correlated movements. Given a new planning CT, DAM estimates the joint distribution over the variables, with each sample from the distribution corresponding to a different deformation. We train our model using dataset of 312 CT pairs with prostate, bladder, and rectum delineations from 38 prostate cancer patients. For 2 additional patients (22 CTs), we compute the contour overlap between real and generated images, and compare the sampled and 'ground truth' distributions of volume and center of mass changes.Results. With a DICE score of 0.86 ± 0.05 and a distance between prostate contours of 1.09 ± 0.93 mm, DAM matches and improves upon previously published PCA-based models, using as few as 8 latent variables. The overlap between distributions further indicates that DAM's sampled movements match the range and frequency of clinically observed daily changes on repeat CTs.Significance. Conditioned only on planning CT values and organ contours of a new patient without any pre-processing, DAM can accurately deformations seen during following treatment sessions, enabling anatomically robust treatment planning and robustness evaluation against inter-fraction anatomical changes.