Potential margin reduction in prostate cancer proton therapy with prompt gamma imaging for online treatment verification.

The potential of proton therapy is currently limited due to large safety margins. We estimated the potential reduction of clinical margins when using prompt gamma imaging (PGI) for online treatment verification of prostate cancer. For two adaptive scenarios a potential reduction relative to clinical practice was evaluated. The use of a trolley-mounted PGI system for online treatment verification to trigger an adaptation reduced the current range margins from 7 mm to 3 mm. In a case example, the dose reduction due to reduced range margins was substantially larger compared to reduced setup margins when using pre-treatment volumetric imaging.

Detectability of Anatomical Changes With Prompt-Gamma Imaging: First Systematic Evaluation of Clinical Application During Prostate-Cancer Proton Therapy.

PURPOSE: The development of online-adaptive proton therapy (PT) is essential to overcome limitations encountered by day-to-day variations of the patient's anatomy. Range verification could play an essential role in an online feedback loop for the detection of treatment deviations such as anatomical changes. Here, we present the results of the first systematic patient study regarding the detectability of anatomical changes by a prompt-gamma imaging (PGI) slit-camera system. METHODS AND MATERIALS: For 15 patients with prostate cancer, PGI measurements were performed during 105 fractions (201 fields) with in-room control computed tomography (CT)acquisitions. Field-wise doses on control CT scans were manually classified as whether showing relevant or non-relevant anatomical changes. This manual classification of the treatment fields was then used to establish an automatic field-wise ground truth based on spot-wise dosimetric range shifts, which were retrieved from integrated depth-dose (IDD) profiles. To determine the detection capability of anatomical changes with PGI, spot-wise PGI-based range shifts were initially compared with corresponding dosimetric IDD range shifts. As final endpoint, the agreement of a developed field-wise PGI classification model with the field-wise ground truth was determined. Therefore, the PGI model was optimized and tested for a subcohort of 131 and 70 treatment fields, respectively. RESULTS: The correlation between PGI and IDD range shifts was high, ρpearson = 0.67 (p < 0.01). Field-wise binary PGI classification resulted in an area under the curve of 0.72 and 0.80 for training and test cohorts, respectively. The model detected relevant anatomical changes in the independent test cohort, with a sensitivity and specificity of 74% and 79%, respectively. CONCLUSIONS: For the first time, evidence of the detection capability of anatomical changes in prostate-cancer PT from clinically acquired PGI data is shown. This emphasizes the benefit of PGI-based range verification and demonstrates its potential for online-adaptive PT.

Automatic detection and classification of treatment deviations in proton therapy from realistically simulated prompt gamma imaging data.

BACKGROUND: A clinical study regarding the potential of range verification in proton therapy (PT) by prompt gamma imaging (PGI) is carried out at our institution. Manual interpretation of the detected spot-wise range shift information is time-consuming, highly complex, and therefore not feasible in a broad routine application. PURPOSE: Here, we present an approach to automatically detect and classify treatment deviations in realistically simulated PGI data for head-and-neck cancer (HNC) treatments using convolutional neural networks (CNNs) and conventional machine learning (ML) approaches. METHODS: For 12 HNC patients and 1 anthropomorphic head phantom (n = 13), pencil beam scanning (PBS) treatment plans were generated, and 1 field per plan was assumed to be monitored with a PGI slit camera system. In total, 386 scenarios resembling different relevant or non-relevant treatment deviations were simulated on planning and control CTs and manually classified into 7 classes: non-relevant changes (NR) and relevant changes (RE) triggering treatment intervention due to range prediction errors (±RP), setup errors in beam direction (±SE), anatomical changes (AC), or a combination of such errors (CB). PBS spots with reliable PGI information were considered with their nominal Bragg peak position for the generation of two 3D spatial maps of 16 × 16 × 16 voxels containing PGI-determined range shift and proton number information. Three complexity levels of simulated PGI data were investigated: (I) optimal PGI data, (II) realistic PGI data with simulated Poisson noise based on the locally delivered proton number, and (III) realistic PGI data with an additional positioning uncertainty of the slit camera following an experimentally determined distribution. For each complexity level, 3D-CNNs were trained on a data subset (n = 9) using patient-wise leave-one-out cross-validation and tested on an independent test cohort (n = 4). Both the binary task of detecting RE and the multi-class task of classifying the underlying error source were investigated. Similarly, four different conventional ML classifiers (logistic regression, multilayer perceptron, random forest, and support vector machine) were trained using five previously established handcrafted features extracted from the PGI data and used for performance comparison. RESULTS: On the test data, the CNN ensemble achieved a binary accuracy of 0.95, 0.96, and 0.93 and a multi-class accuracy of 0.83, 0.81, and 0.76 for the complexity levels (I), (II), and (III), respectively. In the case of binary classification, the CNN ensemble detected treatment deviations in the most realistic scenario with a sensitivity of 0.95 and a specificity of 0.88. The best performing ML classifiers showed a similar test performance. CONCLUSIONS: This study demonstrates that CNNs can reliably detect relevant changes in realistically simulated PGI data and classify most of the underlying sources of treatment deviations. The CNNs extracted meaningful features from the PGI data with a performance comparable to ML classifiers trained on previously established handcrafted features. These results highlight the potential of a reliable, automatic interpretation of PGI data for treatment verification, which is highly desired for a broad clinical application and a prerequisite for the inclusion of PGI in an automated feedback loop for online adaptive PT.

First-In-Human Validation of CT-Based Proton Range Prediction Using Prompt Gamma Imaging in Prostate Cancer Treatments.

PURPOSE: Uncertainty in computed tomography (CT)-based range prediction substantially impairs the accuracy of proton therapy. Direct determination of the stopping-power ratio (SPR) from dual-energy CT (DECT) has been proposed (DirectSPR), and initial validation studies in phantoms and biological tissues have proven a high accuracy. However, a thorough validation of range prediction in patients has not yet been achieved by any means. Here, we present the first systematic validation of CT-based proton range prediction in patients using prompt gamma imaging (PGI). METHODS AND MATERIALS: A PGI slit camera system with improved positioning accuracy, using a floor-based docking station, was used. Its overall uncertainty for range prediction validation was determined experimentally with both x-ray and beam measurements. The accuracy of range prediction in patients was determined from clinical PGI measurements during hypofractionated treatment of 5 patients with prostate cancer - in total 30 fractions with in-room control-CTs. For each pencil-beam-scanning spot, the range shift was obtained by comparing the PGI measurement to a control-CT-based PGI simulation. Three different SPR prediction approaches were applied in simulations: a standard CT-number-to-SPR conversion (Hounsfield look-up table [HLUT]), an adapted HLUT (DECT optimized), and DirectSPR. The spot-wise weighted mean range shift from all spots served as a measure for the accuracy of the respective range prediction approach. RESULTS: A mean range prediction accuracy of 0.0% ± 0.5%, 0.3% ± 0.4%, and 1.8% ± 0.4% was obtained for DirectSPR, adapted HLUT, and standard HLUT, respectively. The overall validation uncertainty of the second-generation PGI slit camera is about 1 mm (2σ) for all approaches, which is smaller than the range prediction uncertainty for deep-seated tumors. CONCLUSIONS: For the first time, range prediction accuracy was assessed in clinical routine using PGI range verification in prostate cancer treatments. Both DECT-derived range prediction approaches agree well with the measured proton range from PGI verification, whereas the standard HLUT approach differs relevantly. These results endorse the recent reduction of clinical safety margins in DirectSPR-based treatment planning in our institution.

Classification of the source of treatment deviation in proton therapy using prompt-gamma imaging information.

PURPOSE: Prompt-gamma imaging (PGI)-based range verification has been successfully implemented in clinical proton therapy recently and its sensitivity to detect treatment deviations is currently investigated. The cause of treatment deviations can be multiple - for example, computed tomography (CT)-based range prediction, patient setup, and anatomical changes. Hence, it would be beneficial, if PGI-based verification would not only detect a treatment deviation but would also be able to directly identify its most probable source. Here, we propose a heuristically derived decision tree approach to automatically classify the sources of range deviation in proton pencil-beam scanning (PBS) treatments of head and neck tumors based on range information obtained with a PGI slit camera. MATERIALS AND METHODS: The decision tree model was iteratively generated on a training dataset of different anatomical complexities, for an anthropomorphic head phantom and patient CT data (planning and control CTs) from five patients. Different range prediction errors, setup changes and relevant and nonrelevant anatomical changes were introduced or derived from control CTs, summing up to a total of 98 training scenarios. Independent validation was performed for another 98 scenarios, derived solely from patient CT data of another seven patients. PBS head and neck treatment plans were generated for the nominal scenario. For all PBS spots in the investigated field, PGI profiles were simulated using a dedicated analytical model of the slit camera for the nominal as well as the different error scenarios. From comparison of PGI profiles for nominal and error scenarios, a spot-wise range shift after spot aggregation with a kernel of 7 mm sigma was determined for each error scenario. The heuristic approach includes a prefiltering of the most suitable PBS spots for PGI treatment verification. From the validation, the accuracy, sensitivity, and specificity of the model were determined. RESULTS: A five-step consecutive filter was developed to preselect PBS spots. On average, 25% of spots (1044 spots) remained as input for the classification model. The derived heuristic decision tree model is based on five parameters: The coefficient of determination (R2 ), the slope and intercept of the linear regression between PGI-derived range shifts and the respectively predicted proton ranges for the investigated PBS spots, as well as the average and standard deviation of the PGI-derived shifts. With this approach, 94 of 98 error scenarios could be classified correctly in validation (accuracy of 96%). A sensitivity and specificity of 100% and 86% were reached. CONCLUSIONS: In this simulation study it was demonstrated that the source of a treatment deviation can be identified from simulated noiseless PGI information in head and neck tumor treatments with high sensitivity and specificity. The application, refinement, and evaluation of the approach on measured PGI data will be the next step to show the clinical feasibility of PGI-based error source classification.

Processing of prompt gamma-ray timing data for proton range measurements at a clinical beam delivery.

In proton therapy, patients benefit from the precise deposition of the dose in the tumor volume due to the interaction of charged particles with matter. Currently, the determination of the beam range in the patient's body during the treatment is not a clinical standard. This lack causes broad safety margins around the tumor, which limits the potential of proton therapy. To overcome this obstacle, different methods are under investigation aiming at the verification of the proton range in real time during the irradiation. One approach is the prompt gamma-ray timing (PGT) method, where the range of the primary protons is derived from time-resolved profiles (PGT spectra) of promptly emitted gamma rays, which are produced along the particle track in tissue. After verifying this novel technique in an experimental environment but far away from treatment conditions, the translation of PGT into clinical practice is intended. Therefore, new hardware was extensively tested and characterized using short irradiation times of 70 ms and clinical beam currents of 2 nA. Experiments were carried out in the treatment room of the University Proton Therapy Dresden. A pencil beam scanning plan was delivered to a target without and with cylindrical air cavities of down to 5 mm thickness. The range shifts of the proton beam induced due to the material variation could be identified from the corresponding PGT spectra, comprising events collected during the delivery of a whole energy layer. Additionally, an assignment of the PGT data to the individual pencil beam spots allowed a spot-wise analysis of the variation of the PGT distribution mean and width, corresponding to range shifts produced by the different air cavities. Furthermore, the paper presents a comprehensive software framework which standardizes future PGT analysis methods and correction algorithms for technical limitations that have been encountered in the presented experiments.