Deep learning prediction of proton and photon dose distributions for paediatric abdominal tumours.

OBJECTIVE: Dose prediction using deep learning networks prior to radiotherapy might lead tomore efficient modality selections. The study goal was to predict proton and photon dose distributions based on the patient-specific anatomy and to assess their clinical usage for paediatric abdominal tumours. MATERIAL AND METHODS: Data from 80 patients with neuroblastoma or Wilms' tumour was included. Pencil beam scanning (PBS) (5 mm/ 3%) and volumetric-modulated arc therapy (VMAT) plans (5 mm) were robustly optimized on the internal target volume (ITV). Separate 3-dimensional patch-based U-net networks were trained to predict PBS and VMAT dose distributions. Doses, planning-computed tomography images and relevant optimization masks (ITV, vertebra and organs-at-risk) of 60 patients were used for training with a 5-fold cross validation. The networks' performance was evaluated by computing the relative error between planned and predicted dose-volume histogram (DVH) parameters for 20 inference patients. In addition, the organs-at-risk mean dose difference between modalities was calculated using planned and predicted dose distributions (ΔDmean = DVMAT-DPBS). Two radiation oncologists performed a blind PBS/VMAT modality selection based on either planned or predicted ΔDmean. RESULTS: Average DVH differences between planned and predicted dose distributions were ≤ |6%| for both modalities. The networks classified the organs-at-risk Dmean difference as a gain (ΔDmean > 0) with 98% precision. An identical modality selection based on planned compared to predicted ΔDmean was made for 18/20 patients. CONCLUSION: Deep learning networks for accurate prediction of proton and photon dose distributions for abdominal paediatric tumours were established. These networks allowing fast dose visualisation might aid in identifying the optimal radiotherapy technique when experience and/or resources are unavailable.

Assessment of range uncertainty in lung-like tissue using a porcine lung phantom and proton radiography.

Thoracic tumours are increasingly considered indications for pencil beam scanned proton therapy (PBS-PT) treatments. Conservative robustness settings have been suggested due to potential range straggling effects caused by the lung micro-structure. Using proton radiography (PR) and a 4D porcine lung phantom, we experimentally assess range errors to be considered in robust treatment planning for thoracic indications. A human-chest-size 4D phantom hosting inflatable porcine lungs and corresponding 4D computed tomography (4DCT) were used. Five PR frames were planned to intersect the phantom at various positions. Integral depth-dose curves (IDDs) per proton spot were measured using a multi-layer ionisation chamber (MLIC). Each PR frame consisted of 81 spots with an assigned energy of 210 MeV (full width at half maximum (FWHM) 8.2 mm). Each frame was delivered five times while simultaneously acquiring the breathing signal of the 4D phantom, using an ANZAI load cell. The synchronised ANZAI and delivery log file information was used to retrospectively sort spots into their corresponding breathing phase. Based on this information, IDDs were simulated by the treatment planning system (TPS) Monte Carlo dose engine on a dose grid of 1 mm. In addition to the time-resolved TPS calculations on the 4DCT phases, IDDs were calculated on the average CT. Measured IDDs were compared with simulated ones, calculating the range error for each individual spot. In total, 2025 proton spots were individually measured and analysed. The range error of a specific spot is reported relative to its water equivalent path length (WEPL). The mean relative range error was 1.2% (1.5 SD 2.3 %) for the comparison with the time-resolved TPS calculations, and 1.0% (1.5 SD 2.2 %) when comparing to TPS calculations on the average CT. The determined mean relative range errors justify the use of 3% range uncertainty for robust treatment planning in a clinical setting for thoracic indications.

Validation of the proton range accuracy and optimization of CT calibration curves utilizing range probing.

Proton therapy is affected by range uncertainty, which is partly caused by an ambiguous conversion from x-ray attenuation to proton stopping power. CT calibration curves, or Hounsfield look-up tables (HLUTs), are institution-specific and may be a source of systematic errors in treatment planning. A range probing method to verify, optimize and validate HLUTs for proton treatment is proposed. An initial HLUT was determined according to the stoichiometric approach. For HLUT validation, three types of animal tissue phantoms were prepared: a pig's head, 'thorax' and femur. CT scans of the phantoms were taken and a structure, simulating a water slab, was added on the scan distal to the phantoms to mimic the detector used for integral depth-dose measurements. The CT scans were imported into the TPS to calculate individual pencil beams directed through the phantoms. The phantoms were positioned at the therapy system isocenter using x-ray imaging. Shoot-through pencil beams were delivered, and depth-dose profiles were measured using a multi-layer ionization chamber. Measured depth-dose curves were compared to the calculated curves and the range error per spot was determined. Based on the water equivalent path length (WEPL) of individual spot, a range error margin was defined. Ratios between measured error and theoretical margin were calculated per spot. The HLUT optimization was performed by identifying systematic shifts of the mean range error per phantom and minimizing the ratios between range errors and uncertainty margins. After optimization, the ratios of the actual range error and the uncertainty margin over the complete data set did not exceed 0.75 (1.5 SD), indicating that the actual errors are covered by the theoretical uncertainty recipe. The feasibility of using range probing to assess range errors was demonstrated. The theoretical uncertainty margins in the institution-specific setting potentially may be reduced by ~25%.

An automated, quantitative, and case-specific evaluation of deformable image registration in computed tomography images.

A prerequisite for adaptive dose-tracking in radiotherapy is the assessment of the deformable image registration (DIR) quality. In this work, various metrics that quantify DIR uncertainties are investigated using realistic deformation fields of 26 head and neck and 12 lung cancer patients. Metrics related to the physiologically feasibility (the Jacobian determinant, harmonic energy (HE), and octahedral shear strain (OSS)) and numerically robustness of the deformation (the inverse consistency error (ICE), transitivity error (TE), and distance discordance metric (DDM)) were investigated. The deformable registrations were performed using a B-spline transformation model. The DIR error metrics were log-transformed and correlated (Pearson) against the log-transformed ground-truth error on a voxel level. Correlations of r ⩾ 0.5 were found for the DDM and HE. Given a DIR tolerance threshold of 2.0 mm and a negative predictive value of 0.90, the DDM and HE thresholds were 0.49 mm and 0.014, respectively. In conclusion, the log-transformed DDM and HE can be used to identify voxels at risk for large DIR errors with a large negative predictive value. The HE and/or DDM can therefore be used to perform automated quality assurance of each CT-based DIR for head and neck and lung cancer patients.

Evaluation of a multi-atlas CT synthesis approach for MRI-only radiotherapy treatment planning.

BACKGROUND AND PURPOSE: Computed tomography (CT) imaging is the current gold standard for radiotherapy treatment planning (RTP). The establishment of a magnetic resonance imaging (MRI) only RTP workflow requires the generation of a synthetic CT (sCT) for dose calculation. This study evaluates the feasibility of using a multi-atlas sCT synthesis approach (sCTa) for head and neck and prostate patients. MATERIAL AND METHODS: The multi-atlas method was based on pairs of non-rigidly aligned MR and CT images. The sCTa was obtained by registering the MRI atlases to the patient's MRI and by fusing the mapped atlases according to morphological similarity to the patient. For comparison, a bulk density assignment approach (sCTbda) was also evaluated. The sCTbda was obtained by assigning density values to MRI tissue classes (air, bone and soft-tissue). After evaluating the synthesis accuracy of the sCTs (mean absolute error), sCT-based delineations were geometrically compared to the CT-based delineations. Clinical plans were re-calculated on both sCTs and a dose-volume histogram and a gamma analysis was performed using the CT dose as ground truth. RESULTS: Results showed that both sCTs were suitable to perform clinical dose calculations with mean dose differences less than 1% for both the planning target volume and the organs at risk. However, only the sCTa provided an accurate and automatic delineation of bone. CONCLUSIONS: Combining MR delineations with our multi-atlas CT synthesis method could enable MRI-only treatment planning and thus improve the dosimetric and geometric accuracy of the treatment, and reduce the number of imaging procedures.