Uncertainty-aware MR-based CT synthesis for robust proton therapy planning of brain tumour.

BACKGROUND AND PURPOSE: Deep learning techniques excel in MR-based CT synthesis, but missing uncertainty prediction limits its clinical use in proton therapy. We developed an uncertainty-aware framework and evaluated its efficiency in robust proton planning. MATERIALS AND METHODS: A conditional generative-adversarial network was trained on 64 brain tumour patients with paired MR-CT images to generate synthetic CTs (sCT) from combined T1-T2 MRs of three orthogonal planes. A Bayesian neural network predicts Laplacian distributions for all voxels with parameters (µ, b). A robust proton plan was optimized using three sCTs of µ and µ±b. The dosimetric differences between the plan from sCT (sPlan) and the recalculated plan (rPlan) on planning CT (pCT) were quantified for each patient. The uncertainty-aware robust plan was compared to conventional robust (global ± 3 %) and non-robust plans. RESULTS: In 8-fold cross-validation, sCT-pCT image differences (Mean-Absolute-Error) were 80.84 ± 9.84HU (body), 35.78 ± 6.07HU (soft tissues) and 221.88 ± 31.69HU (bones), with Dice scores of 90.33 ± 2.43 %, 95.13 ± 0.80 %, and 85.53 ± 4.16 %, respectively. The uncertainty distribution positively correlated with absolute prediction error (Correlation Coefficient: 0.62 ± 0.01). The uncertainty-conditioned robust optimisation improved the rPlan-sPlan agreement, e.g., D95 absolute difference (CTV) was 1.10 ± 1.24 % compared to conventional (1.64 ± 2.71 %) and non-robust (2.08 ± 2.96 %) optimisation. This trend was consistent across all target and organs-at-risk indexes. CONCLUSION: The enhanced framework incorporates 3D uncertainty prediction and generates high-quality sCTs from MR images. The framework also facilitates conditioned robust optimisation, bolstering proton plan robustness against network prediction errors. The innovative feature of uncertainty visualisation and robust analyses contribute to evaluating sCT clinical utility for individual patients.

A fast analytical dose calculation approach for MRI-guided proton therapy.

Objective.Magnetic resonance (MR) is an innovative technology for online image guidance in conventional radiotherapy and is also starting to be considered for proton therapy as well. For MR-guided therapy, particularly for online plan adaptations, fast dose calculation is essential. Monte Carlo (MC) simulations, however, which are considered the gold standard for proton dose calculations, are very time-consuming. To address the need for an efficient dose calculation approach for MRI-guided proton therapy, we have developed a fast GPU-based modification of an analytical dose calculation algorithm incorporating beam deflections caused by magnetic fields.Approach.Proton beams (70-229 MeV) in orthogonal magnetic fields (0.5/1.5 T) were simulated using TOPAS-MC and central beam trajectories were extracted to generate look-up tables (LUTs) of incremental rotation angles as a function of water-equivalent depth. Beam trajectories are then reconstructed using these LUTs for the modified ray casting dose calculation. The algorithm was validated against MC in water, different materials and for four example patient cases, whereby it has also been fully incorporated into a treatment plan optimisation regime.Main results.Excellent agreement between analytical and MC dose distributions could be observed with sub-millimetre range deviations and differences in lateral shifts <2 mm even for high densities (1000 HU). 2%/2 mm gamma pass rates were comparable to the 0 T scenario and above 94.5% apart for the lung case. Further, comparable treatment plan quality could be achieved regardless of magnetic field strength.Significance.A new method for accurate and fast proton dose calculation in magnetic fields has been developed and successfully implemented for treatment plan optimisation.

Intensity modulated proton therapy plan generation in under ten seconds.

Background: Treatment planning for intensity modulated proton therapy (IMPT) can be significantly improved by reducing the time for plan calculation, facilitating efficient sampling of the large solution space characteristic of IMPT treatments. Additionally, fast plan generation is a key for online adaptive treatments, where the adapted plan needs to be ideally available in a few seconds. However, plan generation is a computationally demanding task and, although dose restoration methods for adaptive therapy have been proposed, computation times remain problematic. Material and methods: IMPT plan generation times were reduced by the development of dedicated graphical processing unit (GPU) kernels for our in-house, clinically validated, dose and optimization algorithms. The kernels were implemented into a coherent system, which performed all steps required for a complete treatment plan generation. Results: Using a single GPU, our fast implementation was able to generate a complete new treatment plan in 5-10 sec for typical IMPT cases, and in under 25 sec for plans to very large volumes such as for cranio-spinal axis irradiations. Although these times did not include the manual input of optimization parameters or a final clinical dose calculation, they included all required computational steps, including reading of CT and beam data. In addition, no compromise was made on plan quality. Target coverage and homogeneity for four patient plans improved (by up to 6%) or remained the same (changes <1%). No worsening of dose-volume parameters of the relevant organs at risk by more than 0.5% was observed. Conclusions: Fast plan generation with a clinically validated dose calculation and optimizer is a promising approach for daily adaptive proton therapy, as well as for automated or highly interactive planning.

Evaluation of the ray-casting analytical algorithm for pencil beam scanning proton therapy.

For pencil beam scanned (PBS) proton therapy, analytical dose calculation engines are still typically used for the optimisation process, and often for the final evaluation of the plan. Recently however, the suitability of analytical calculations for planning PBS treatments has been questioned. Conceptually, the two main approaches for these analytical dose calculations are the ray-casting (RC) and the pencil-beam (PB) method. In this study, we compare dose distributions and dosimetric indices, calculated on both the clinical dose calculation grid and as a function of dose grid resolution, to Monte Carlo (MC) calculations. The analysis is done using a comprehensive set of clinical plans which represent a wide choice of treatment sites. When analysing dose difference histograms for relative treatment plans, pencil beam calculations with double grid resolution perform best, with on average 97.7%/91.9% (RC), 97.9%/92.7% (RC, double grid resolution), 97.6%/91.0% (PB) and 98.6%/94.0% (PB, double grid resolution) of voxels agreeing within ±5%/± 3% between the analytical and the MC calculations. Even though these point-to-point dose comparison shows differences between analytical and MC calculations, for all algorithms, clinically relevant dosimetric indices agree within ±4% for the PTV and within ±5% for critical organs. While the clinical agreement depends on the treatment site, there is no substantial difference of indices between the different algorithms. The pencil-beam approach however comes at a higher computational cost than the ray-casting calculation. In conclusion, we would recommend using the ray-casting algorithm for fast dose optimization and subsequently combine it with one MC calculation to scale the absolute dose and assure the quality of the treatment plan.

Log file based Monte Carlo calculations for proton pencil beam scanning therapy.

Patient specific quality assurance is crucial to guarantee safety in proton pencil beam scanning. In current clinical practice, this requires extensive, time consuming measurements. Additionally, these measurements do not consider the influence of density heterogeneities in the patient and are insensitive to delivery errors. In this work, we investigate the use of log file based Monte Carlo calculations for dose reconstructions in the patient CT, which takes the combined influence of calculational and delivery errors into account. For one example field, 87%/90% of the voxels agree within ±3% when taking either calculational or delivery uncertainties into account (analytical versus Monte Carlo calculation/Monte Carlo from planned versus Monte Carlo from log file). 78% agree when considering both uncertainties simultaneously (nominal field versus Monte Carlo from log files). We then show the application of the log file based Monte Carlo calculations as a patient specific quality assurance tool for a set of five patients (16 fields) treated for different indications. For all fields, absolute dose scaling factors based on the log file Monte Carlo agree within ±3% to the measurement based absolute dose scaling. Relative comparison shows that more than 90% of the voxels agree within ± 5% between the analytical calculated plan and the Monte Carlo based on log files. The log file based Monte Carlo approach is an end-to-end test incorporating all requirements of patient specific quality assurance. It has the potential to reduce the workload and therefore to increase the patient throughput, while simultaneously enabling more accurate dose verification directly in the patient geometry.

Contour scanning, multi-leaf collimation and the combination thereof for proton pencil beam scanning.

In proton therapy, the lateral fall-off is often used to spare critical organs. It is therefore crucial to improve the penumbra for proton pencil beam scanning. However, previous work has shown that collimation may not be necessary for depths of >15 cm in water. As such, in this work we investigate the effectiveness of a thin multi leaf collimator (just thick enough to completely stop protons with ranges of <15 cm in water) for energy layer specific collimation in patient geometries, when applied in combination with both grid and contour scanned PBS proton therapy. For this, an analytical model of collimated beam shapes, based solely on data available in the treatment planning system, has been included in the optimization, with the resulting optimised plans then being recalculated using Monte Carlo in order to most accurately simulate the full physics effects of the collimator. For grid based scanning, energy specific collimation has been found to reduce the V30 outside the PTV by 19.8% for an example patient when compared to the same pencil beam placement without collimation. V30 could be even reduced by a further 5.6% when combining collimation and contour scanning. In addition, mixed plans, consisting of contour scanning for deep fields (max range >15 cm WER) and collimated contour scanning for superficial fields (<15 cm), have been created for four patients, by which V30 could be reduced by 0.8% to 8.0% and the mean dose to the brain stem by 1.5% to 3.3%. Target dose homogeneity however is not substantially different when compared to the best un-collimated scenario. In conclusion, we demonstrate the potential advantages of a thin, multi leaf collimator in combination with contour scanning for energy layer specific collimation in PBS proton therapy.

A novel dose-volume metric for optimizing therapeutic ratio through fractionation: retrospective analysis of lung cancer treatments.

PURPOSE: To explore the potential of a novel dose-volume based metric to assist in the selection of optimal fractionation schedules for lung cancer patients. METHODS: Selecting the dose per fraction that maximizes the therapeutic ratio via a linear-quadratic effect on normal tissue complication probability and tumor cell survival is an optimization problem. The mathematical solution reveals that the optimal fractionation schedule is determined by a generalized dose ratio between the normal tissue and the tumor, here termed the bifurcation number B, that can be derived from the dose-volume histogram of the normal tissue. The bifurcation number characterizes the volume effect of a normal tissue and its dependency on the fractionation schedule. The clinical relevance of the bifurcation number was evaluated in 46 patients previously treated for nonsmall cell lung cancer (NSCLC) according to various fractionation protocols. Bifurcation numbers were computed for both lung and esophagus as the normal tissues. RESULTS: The value of the bifurcation number determines whether the volume effect reverses the traditional radiobiological advantage of small dose per fraction for the normal tissue. If B is smaller than the ratio of alpha/beta ratios between normal tissue and tumor, then a single fraction is optimal; otherwise the optimal treatment is an infinite number of doses (hence the name "bifurcation" number). These fractionation schedules correspond clinically to hypo- and standard/hyperfractionation, respectively. Compared with traditional dose-volume metrics, the bifurcation number is a unitless ratio and independent of dose fractionation. The B-numbers derived from the clinical treatment plans are also strongly consistent with historically prescribed clinical fractionation protocols for NSCLC treatments. The B-numbers for esophagus and lung for all patients receiving a high dose per fraction protocol (>7.5 Gy/fraction) were all smaller than the B-numbers for the patients receiving standard 2 Gy/fraction, with the numbers for the 3 Gy/fraction group in between. CONCLUSIONS: The bifurcation numbers are strongly consistent with prescribed clinical fractionation protocols for NSCLC treatments. Due to their scale-free property the B-numbers may assist in the selection of an appropriate fractionation once the dose distribution has been optimized.