A deep LSTM autoencoder-based framework for predictive maintenance of a proton radiotherapy delivery system.

INTRODUCTION: Unscheduled machine downtime can cause treatment interruptions and adversely impact patient treatment outcomes. Conventional Quality Assurance (QA) programs of a proton Pencil Beam Scanning (PBS) system ensure its operational performance by keeping the beam parameters within clinical tolerances but often do not reveal the underlying issues of the device prior to a machine malfunction event. In this study, we propose a Predictive Maintenance (PdM) approach that leverages an advanced analytical tool built on a deep neural network to detect treatment delivery machine issues early. METHODS: Beam delivery log file data from daily QA performed at the Burr Proton Center of Massachusetts General Hospital were collected. A novel PdM framework consisting of long short-term memory-based autoencoder (LSTM-AE) modeling of the proton PBS delivery system and a Mahalanobis distance-based error metric evaluation was constructed to detect rare anomalous machine events. These included QA beam pauses, clinical operational issues, and treatment interruptions. The model was trained in an unsupervised fashion on the QA data of normal sessions so that the model learned characteristics of normal machine operation. The anomaly is quantified as the multivariate deviation between the model predicted data and the measured data of the day using Mahalanobis distance (M-Score). Two-layer and three-layer Long short-term memory-based stacked autoencoder (LSTM-SAE) models were optimized for exploring model performance improvement. Model validation was performed with two clinical datasets and was analyzed using the area under the precision-recall curve (AUPRC) and the area under the receiver operating characteristic (AUROC). RESULTS: LSTM-SAE models showed strong performance in predicting QA beam pauses for both clinical validation datasets. Despite severe skew in the dataset, the model achieved AUPRC of 0.60 and 0.82 and AUROC of 0.75 and 0.92 in the respective 2018 and 2020 datasets. Moreover, these amount to 2.8-fold and 10.7-fold enhancement compared to the respective baseline event rates. In addition, in terms of treatment interruption events, model prediction enabled 3.88-fold and 51.2-fold detection improvement, while the detection improvement for clinical operational issues was 1.04-fold and 1.37-fold, respectively, in the 2018 and 2020 datasets. CONCLUSION: Our novel deep LSTM-SAE-based framework allows for highly discriminative prediction of anomalous machine events and demonstrates great promise for enabling PdM for proton PBS beam delivery.

Iterative optimization of relative stopping power by single detector based multi-projection proton radiography.

In proton therapy, range uncertainties induced by the conversion from x-ray CT (xCT) Hounsfield units (HU) to relative stopping power (RSP) compromise the precision of dose delivery. To reduce range uncertainties induced by HU-converted RSPs, this study investigates optimizing the RSP of individual voxels in xCT iteratively based on multi-projection proton radiography (pRG) acquired using a single amorphous silicon flat panel imager. Time-resolved dose rate functions (DRF) were measured by the imager placed downstream of a test phantom consisting of tissue substitute materials. Water equivalent path lengths (WEPL) in the pRG were derived. By rotating the phantom, multiple pRG projections were acquired at angles from 0 to 358° with an increment of 2°. X-ray CT of the phantom was acquired and co-registered with the pRG acquisition coordinates. RSPs of individual xCT voxels were optimized iteratively by minimizing the difference between the measured WEPLs and the calculated WEPLs by ray tracing with HU-converted RSPs. Pixels in pRGs that exhibited severe proton range mixing were rejected for the optimization. Tikhonov regularization was applied under the assumption that HU-converted RSPs are inaccurate, but the inaccuracy is within a few percent. While ~50% of WEPL pixels were rejected due to severe range mixing in pRG, RSPs of >90% CT voxels could still be optimized if multiple pRG projections, e.g. ⩾12, around the phantom are utilized. For tissue substitute materials in a cylindrical phantom, percentage errors of RSPs were reduced from a range of -8% to +4% to be within ±2%. Further optimization, achieved by implementing a material-specific regularization parameter, reduced percent errors to be within ±0.5%. This study demonstrates the concept of optimizing RSPs of individual CT voxels with multi-projection pRGs acquired by a single flat panel imager, which could be further explored and implemented in proton therapy to reduce range uncertainties.

New horizons in particle therapy systems.

Particle therapy is rapidly expanding and claiming its position as the treatment modality of choice in teletherapy. However, the rate of expansion continues to be restricted by the size and cost of the associated particle therapy systems and their operation. Additional technical limitations such as dose delivery rate, treatment process efficiency, and achievement of superior dose conformity potentially hinder the complete fulfillment of the promise of particle therapy. These topics are explored in this review considering the current state of particle therapy systems and what improvements are required to overcome the current challenges. Beam production systems (accelerators), beam transport systems including gantries and beam delivery systems are addressed explicitly in these regards.

Improvement of single detector proton radiography by incorporating intensity of time-resolved dose rate functions.

Proton radiography, which images patients with the same type of particles as those with which they are to be treated, is a promising approach to image guidance and water equivalent path length (WEPL) verification in proton radiation therapy. We have shown recently that proton radiographs could be obtained by measuring time-resolved dose rate functions (DRFs) using an x-ray amorphous silicon flat panel. The WEPL values were derived solely from the root-mean-square (RMS) of DRFs, while the intensity information in the DRFs was filtered out. In this work, we explored the use of such intensity information for potential improvement in WEPL accuracy and imaging quality. Three WEPL derivation methods based on, respectively, the RMS only, the intensity only, and the intensity-weighted RMS were tested and compared in terms of the quality of obtained radiograph images and the accuracy of WEPL values. A Gammex CT calibration phantom containing inserts made of various tissue substitute materials with independently measured relative stopping powers (RSP) was used to assess the imaging performances. Improved image quality with enhanced interfaces was achieved while preserving the accuracy by using intensity information in the calibration. Other objects, including an anthropomorphic head phantom, a proton therapy range compensator, a frozen lamb's head and an 'image quality phantom' were also imaged. Both the RMS only and the intensity-weighted RMS methods derived RSPs within ± 1% for most of the Gammex phantom inserts, with a mean absolute percentage error of 0.66% for all inserts. In the case of the insert with a titanium rod, the method based on RMS completely failed, whereas that based on the intensity-weighted RMS was qualitatively valid. The use of intensity greatly enhanced the interfaces between different materials in the obtained WEPL images, suggesting the potential for image guidance in areas such as patient positioning and tumor tracking by proton radiography.

Characterization of proton pencil beam scanning and passive beam using a high spatial resolution solid-state microdosimeter.

PURPOSE: This work aims to characterize a proton pencil beam scanning (PBS) and passive double scattering (DS) systems as well as to measure parameters relevant to the relative biological effectiveness (RBE) of the beam using a silicon on insulator (SOI) microdosimeter with well-defined 3D sensitive volumes (SV). The dose equivalent downstream and laterally outside of a clinical PBS treatment field was assessed and compared to that of a DS beam. METHODS: A novel silicon microdosimeter with well-defined 3D SVs was used in this study. It was connected to low noise electronics, allowing for detection of lineal energies as low as 0.15 keV/µm. The microdosimeter was placed at various depths in a water phantom along the central axis of the proton beam, and at the distal part of the spread-out Bragg peak (SOBP) in 0.5 mm increments. The RBE values of the pristine Bragg peak (BP) and SOBP were derived using the measured microdosimetric lineal energy spectra as inputs to the modified microdosimetric kinetic model (MKM). Geant4 simulations were performed in order to verify the calculated depth-dose distribution from the treatment planning system (TPS) and to compare the simulated dose-mean lineal energy to the experimental results. RESULTS: For a 131 MeV PBS spot (124.6 mm R90 range in water), the measured dose-mean lineal energy yD¯ increased from 2 keV/µm at the entrance to 8 keV/µm in the BP, with a maximum value of 10 keV/µm at the distal edge. The derived RBE distribution for the PBS beam slowly increased from 0.97 ± 0.14 at the entrance to 1.04 ± 0.09 proximal to the BP, then to 1.1 ± 0.08 in the BP, and steeply rose to 1.57 ± 0.19 at the distal part of the BP. The RBE distribution for the DS SOBP beam was approximately 0.96 ± 0.16 to 1.01 ± 0.16 at shallow depths, and 1.01 ± 0.16 to 1.28 ± 0.17 within the SOBP. The RBE significantly increased from 1.29 ± 0.17 to 1.43 ± 0.18 at the distal edge of the SOBP. CONCLUSIONS: The SOI microdosimeter with its well-defined 3D SV has applicability in characterizing proton radiation fields and can measure relevant physical parameters to model the RBE with submillimeter spatial resolution. It has been shown that for a physical dose of 1.82 Gy at the BP, the derived RBE based on the MKM model increased from 1.14 to 1.6 in the BP and its distal part. Good agreement was observed between the experimental and simulation results, confirming the potential application of SOI microdosimeter with 3D SV for quality assurance in proton therapy.

PBS machine interlocks using EWMA.

Delivery of pencil beam scanning (PBS) requires the on-line measurement of several beam parameters. If the measurement is outside of specified tolerances and a binary threshold algorithm is used, the beam will be paused. Given instrumentation and statistical noise such a system can lead to many pauses which could increase the treatment time. Statistical quality control methods are typically used on manufacturing lines to monitor a process and give early detection of a gradual problem and stop the process if a deviation is statistically significant. These methods can be used to develop a more intuitive algorithm for (PBS) delivery systems that is robust and safe and leads to decreased treatment times. The Exponentially Weighted Moving Average (EWMA) control scheme monitors deviations in beam properties which are averaged over a specified number of measurements with greater weight applied to the more recent ones. Simulation of an EWMA-style algorithm safely detected shifts in random and systematic delivery errors without false alarms. Binary and EWMA methods can be combined for improved reliability without sacrificing patient safety. In the EWMA method, the mean of a beam property can be related to systematic uncertainties and the standard deviation can be related to random uncertainties. This method allows one to have separate interlock levels for each type of uncertainty and to detect systematic trends.

Interpolation of tabulated proton Bragg peaks.

Treatment planning databases for pencil beam scanning can be large, difficult to manage and problematic for quality assurance when they contain tabulated Bragg peaks at small range resolution. Smaller range resolution, in the absence of an accurate interpolation method, improves the accuracy in dose calculations. In this work, we derive an approximate scaling function to interpolate between tabulated Bragg peaks, and determine the accuracy of this interpolation technique and the minimum number of tabulated peaks in a treatment planning database. With the new interpolation technique, three tabulated mono-energetic Bragg peaks (N = 3) are a suitable lower limit for N to achieve interpolation accuracy better than ±1% of the maximum dose in pristine and spread out Bragg peaks for ranges between 6.8 and 32.1 cm of water.

Numerical solutions of the γ-index in two and three dimensions.

The γ-index is used routinely to establish correspondence between two dose distributions. The definition of the γ-index can be written with a single equation but solving this equation at millions of points is computationally expensive, especially in three dimensions. Our goal is to extend the vector-equation method in Bakai et al (2003 Phys. Med. Biol.48 3543-53) to higher order for better accuracy and, as important, to determine the magnitude of accuracy in a higher order solution. We construct a numerical framework for calculating the γ-index in two and three dimensions and present an efficient method for calculating the γ-index with zeroth-, first- and second-order methods using tricubic spline interpolation. For an intensity-modulated radiation therapy example with 1.78 × 106 voxels, the zeroth-order, first-order, first-order iterations and semi-second-order methods calculate the three-dimensional γ-index in 1.5, 4.7, 34.7 and 35.6 s with 36.7%, 1.1%, 0.2% and 0.8% accuracy, respectively. The accuracy of linear interpolation with this example is 1.0%. We present efficient numerical methods for calculating the three-dimensional γ-index with tricubic spline interpolation. The first-order method with iterations is the most accurate and fastest choice of the numerical methods if the dose distributions may have large second-order gradients. Furthermore, the difference between iterations can be used to determine the accuracy of the method.

Monte Carlo study of the potential reduction in out-of-field dose using a patient-specific aperture in pencil beam scanning proton therapy.

This study is aimed at identifying the potential benefits of using a patient-specific aperture in proton beam scanning. For this purpose, an accurate Monte Carlo model of the pencil beam scanning (PBS) proton therapy (PT) treatment head at Massachusetts General Hospital (MGH) was developed based on an existing model of the passive double-scattering (DS) system. The Monte Carlo code specifies the treatment head at MGH with sub-millimeter accuracy. The code was configured based on the results of experimental measurements performed at MGH. This model was then used to compare out-of-field doses in simulated DS treatments and PBS treatments. For the conditions explored, the penumbra in PBS is wider than in DS, leading to higher absorbed doses and equivalent doses adjacent to the primary field edge. For lateral distances greater than 10 cm from the field edge, the doses in PBS appear to be lower than those observed for DS. We found that placing a patient-specific aperture at nozzle exit during PBS treatments can potentially reduce doses lateral to the primary radiation field by over an order of magnitude. In conclusion, using a patient-specific aperture has the potential to further improve the normal tissue sparing capabilities of PBS.

Golden beam data for proton pencil-beam scanning.

Proton, as well as other ion, beams applied by electro-magnetic deflection in pencil-beam scanning (PBS) are minimally perturbed and thus can be quantified a priori by their fundamental interactions in a medium. This a priori quantification permits an optimal reduction of characterizing measurements on a particular PBS delivery system. The combination of a priori quantification and measurements will then suffice to fully describe the physical interactions necessary for treatment planning purposes. We consider, for proton beams, these interactions and derive a 'Golden' beam data set. The Golden beam data set quantifies the pristine Bragg peak depth-dose distribution in terms of primary, multiple Coulomb scatter, and secondary, nuclear scatter, components. The set reduces the required measurements on a PBS delivery system to the measurement of energy spread and initial phase space as a function of energy. The depth doses are described in absolute units of Gy(RBE) mm² Gp⁻¹, where Gp equals 109 (giga) protons, thus providing a direct mapping from treatment planning parameters to integrated beam current. We used these Golden beam data on our PBS delivery systems and demonstrated that they yield absolute dosimetry well within clinical tolerance.

A case study in proton pencil-beam scanning delivery.

PURPOSE: We completed an implementation of pencil-beam scanning (PBS), a technology whereby a focused beam of protons, of variable intensity and energy, is scanned over a plane perpendicular to the beam axis and in depth. The aim of radiotherapy is to improve the target to healthy tissue dose differential. We illustrate how PBS achieves this aim in a patient with a bulky tumor. METHODS AND MATERIALS: Our first deployment of PBS uses "broad" pencil-beams ranging from 20 to 35 mm (full-width-half-maximum) over the range interval from 32 to 7 g/cm(2). Such beam-brushes offer a unique opportunity for treating bulky tumors. We present a case study of a large (4,295 cc clinical target volume) retroperitoneal sarcoma treated to 50.4 Gy relative biological effectiveness (RBE) (presurgery) using a course of photons and protons to the clinical target volume and a course of protons to the gross target volume. RESULTS: We describe our system and present the dosimetry for all courses and provide an interdosimetric comparison. DISCUSSION: The use of PBS for bulky targets reduces the complexity of treatment planning and delivery compared with collimated proton fields. In addition, PBS obviates, especially for cases as presented here, the significant cost incurred in the construction of field-specific hardware. PBS offers improved dose distributions, reduced treatment time, and reduced cost of treatment.

PET/CT imaging for treatment verification after proton therapy: a study with plastic phantoms and metallic implants.

The feasibility of off-line positron emission tomography/computed tomography (PET/CT) for routine three dimensional in-vivo treatment verification of proton radiation therapy is currently under investigation at Massachusetts General Hospital in Boston. In preparation for clinical trials, phantom experiments were carried out to investigate the sensitivity and accuracy of the method depending on irradiation and imaging parameters. Furthermore, they addressed the feasibility of PET/CT as a robust verification tool in the presence of metallic implants. These produce x-ray CT artifacts and fluence perturbations which may compromise the accuracy of treatment planning algorithms. Spread-out Bragg peak proton fields were delivered to different phantoms consisting of polymethylmethacrylate (PMMA), PMMA stacked with lung and bone equivalent materials, and PMMA with titanium rods to mimic implants in patients. PET data were acquired in list mode starting within 20 min after irradiation at a commercial luthetium-oxyorthosilicate (LSO)-based PET/CT scanner. The amount and spatial distribution of the measured activity could be well reproduced by calculations based on the GEANT4 and FLUKA Monte Carlo codes. This phantom study supports the potential of millimeter accuracy for range monitoring and lateral field position verification even after low therapeutic dose exposures of 2 Gy, despite the delay between irradiation and imaging. It also indicates the value of PET for treatment verification in the presence of metallic implants, demonstrating a higher sensitivity to fluence perturbations in comparison to a commercial analytical treatment planning system. Finally, it addresses the suitability of LSO-based PET detectors for hadron therapy monitoring. This unconventional application of PET involves countrates which are orders of magnitude lower than in diagnostic tracer imaging, i.e., the signal of interest is comparable to the noise originating from the intrinsic radioactivity of the detector itself. In addition to PET alone, PET/CT imaging provides accurate information on the position of the imaged object and may assess possible anatomical changes during fractionated radiotherapy in clinical applications.

Intra- and interfractional patient motion for a variety of immobilization devices.

The magnitude of inter- and intrafractional patient motion has been assessed for a broad set of immobilization devices. Data was analyzed for the three ordinal directions--left-right (x), sup-inf (y), and ant-post (z)--and the combined spatial displacement. We have defined "rigid" and "non-rigid" immobilization devices depending on whether they could be rigidly and reproducibly connected to the treatment couch or not. The mean spatial displacement for intrafractional motion for rigid devices is 1.3 mm compared to 1.9 mm for nonrigid devices. The modified Gill-Thomas-Cosman frame performed best at controlling intrafractional patient motion, with a 95% probability of observing a three-dimensional (3D) vector length of motion (v95) of less than 1.8 mm, but could not be evaluated for interfractional motion. All other rigid and nonrigid immobilization devices had a v95 of more than 3 mm for intrafractional patient motion. Interfractional patient motion was only evaluated for the rigid devices. The mean total interfractional displacement was at least 3.0 mm for these devices while v95 was at least 6.0 mm.

The prediction of output factors for spread-out proton Bragg peak fields in clinical practice.

The reliable prediction of output factors for spread-out proton Bragg peak (SOBP) fields in clinical practice remained unrealized due to a lack of a consistent theoretical framework and the great number of variables introduced by the mechanical devices necessary for the production of such fields. These limitations necessitated an almost exclusive reliance on manual calibration for individual fields and empirical, ad hoc, models. We recently reported on a theoretical framework for the prediction of output factors for such fields. In this work, we describe the implementation of this framework in our clinical practice. In our practice, we use a treatment delivery nozzle that uses a limited, and constant, set of mechanical devices to produce SOBP fields over the full extent of clinical penetration depths, or ranges, and modulation widths. This use of a limited set of mechanical devices allows us to unfold the physical effects that affect the output factor. We describe these effects and their incorporation into the theoretical framework. We describe the calibration and protocol for SOBP fields, the effects of apertures and range-compensators and the use of output factors in the treatment planning process.