Technical Note: Characterization of clinical linear accelerator triggering latency for motion management system development

Convolutional LSTM model for cine image prediction of abdominal motion

Objective.In this study, we tackle the challenge of latency in magnetic resonance linear accelerator (MR-Linac) systems, which compromises target coverage accuracy in gated real-time radiotherapy. Our focus is on enhancing motion prediction precision in abdominal organs to address this issue. We developed a convolutional long short-term memory (convLSTM) model, utilizing 2D cine magnetic resonance (cine-MR) imaging for this purpose.Approach.Our model, featuring a sequence-to-one architecture with six input frames and one output frame, employs structural similarity index measure (SSIM) as loss function. Data was gathered from 17 cine-MRI datasets using the Philips Ingenia MR-sim system and an Elekta Unity MR-Linac equivalent sequence, focusing on regions of interest (ROIs) like the stomach, liver, pancreas, and kidney. The datasets varied in duration from 1 to 10 min.Main results.The study comprised three main phases: hyperparameter optimization, individual training, and transfer learning with or without fine-tuning. Hyperparameters were initially optimized to construct the most effective model. Then, the model was individually applied to each dataset to predict images four frames ahead (1.24-3.28 s). We evaluated the model's performance using metrics such as SSIM, normalized mean square error, normalized correlation coefficient, and peak signal-to-noise ratio, specifically for ROIs with target motion. The average SSIM values achieved were 0.54, 0.64, 0.77, and 0.66 for the stomach, liver, kidney, and pancreas, respectively. In the transfer learning phase with fine-tuning, the model showed improved SSIM values of 0.69 for the liver and 0.78 for the kidney, compared to 0.64 and 0.37 without fine-tuning.Significance. The study's significant contribution is demonstrating the convLSTM model's ability to accurately predict motion for multiple abdominal organs using a Unity-equivalent MR sequence. This advancement is key in mitigating latency issues in MR-Linac radiotherapy, potentially improving the precision and effectiveness of real-time treatment for abdominal cancers.

Improving respiratory signal prediction with a deep neural network and simple changes to the input and output data format

Objective.To improve respiratory gating accuracy and radiation treatment throughput, we developed a generalized model based on a deep neural network (DNN) for predicting any given patient's respiratory motion.Approach.Our model uses long short-term memory (LSTM) based on a recurrent neural network (RNN), and improves upon common techniques. The first improvement is that the data input is not a one-dimensional sequence, but two-dimensional block data. This shortens the input sequence length, reducing computation time. Second, the output is not a scalar, but a sequence prediction. This increases the amount of available data, allowing improved prediction accuracy. For training and evaluation of our model, 434 sets of real-time position management data were retrospectively collected from clinical studies. The data were separated in a ratio of 4:1, with the larger set used for training models and the remaining set used for testing. We measured the accuracy of respiratory signal prediction and amplitude-based gating with prediction windows equaling 133, 333, and 533 ms. This new model was compared with the original LSTM and a non-recurrent DNN model.Main results.The mean absolute errors with the prediction window at 133, 333 and 533 ms were 0.036, 0.084, 0.119 with our model; 0.049, 0.14, 0.246 with the original LSTM-based model; and 0.041, 0.119, 0.16 with the non-recurrent DNN model, respectively. The computation time were 0.66 ms with our model; 0.63 ms the original LSTM-based model; 1.60 ms the non-recurrent DNN model, respectively. The accuracies of amplitude-based gating with the same prediction window settings and a duty cycle of approximately 50% were 98.3%, 95.8% and 92.7% with our model, 97.6%, 93.9% and 87.2% with the original LSTM-based model; and 97.9%, 94.3% and 89.5% with the non-recurrent DNN model, respectively.Significance.Our RNN algorithm for respiratory signal prediction successfully estimated tumor positions. We believe it will be useful in respiratory signal prediction technology.

A simple method to measure the gating latencies in photon and proton based radiotherapy using a scintillating crystal

BACKGROUND

In respiratory gated radiotherapy, low latency between target motion into and out of the gating window and actual beam-on and beam-off is crucial for the treatment accuracy. However, there is presently a lack of guidelines and accurate methods for gating latency measurements.

PURPOSE

To develop a simple and reliable method for gating latency measurements that work across different radiotherapy platforms.

METHODS

Gating latencies were measured at a Varian ProBeam (protons, RPM gating system) and TrueBeam (photons, TrueBeam gating system) accelerator. A motion-stage performed 1 cm vertical sinusoidal motion of a marker block that was optically tracked by the gating system. An amplitude gating window was set to cover the posterior half of the motion (0-0.5 cm). Gated beams were delivered to a 5 mm cubic scintillating ZnSe:O crystal that emitted visible light when irradiated, thereby directly showing when the beam was on. During gated beam delivery, a video camera acquired images at 120 Hz of the moving marker block and light-emitting crystal. After treatment, the block position and crystal light intensity were determined in all video frames. Two methods were used to determine the gate-on (τ_on ) and gate-off (τ_off ) latencies. By method 1, the video was synchronized with gating log files by temporal alignment of the same block motion recorded in both the video and the log files. τ_on was defined as the time from the block entered the gating window (from gating log files) to the actual beam-on as detected by the crystal light. Similarly, τ_off was the time from the block exited the gating window to beam-off. By method 2, τ_on and τ_off were found from the videos alone using motion of different sine periods (1-10 s). In each video, a sinusoidal fit of the block motion provided the times T_min of the lowest block position. The mid-time, T_mid-light , of each beam-on period was determined as the time halfway between crystal light signal start and end. It can be shown that the directly measurable quantity T_mid-light - T_min  = (τ_off +τ_on )/2, which provided the sum (τ_off +τ_on ) of the two latencies. It can also be shown that the beam-on (i.e., crystal light) duration ΔT_light increases linearly with the sine period and depends on τ_off - τ_on : ΔT_light  = constant•period+(τ_off - τ_on ). Hence, a linear fit of ΔT_light as a function of the period provided the difference of the two latencies. From the sum (τ_off +τ_on ) and difference (τ_off - τ_on ), the individual latencies were determined.

RESULTS

Method 1 resulted in mean (±SD) latencies of τ_on  = 255 ± 33 ms, τ_off  = 82 ± 15 ms for the ProBeam and τ_on  = 84 ± 13 ms, τ_off  = 44 ± 11 ms for the TrueBeam. Method 2 resulted in latencies of τ_on  = 255 ± 23 ms, τ_off  = 95 ± 23 ms for the ProBeam and τ_on  = 83 ± 8 ms, τ_off  = 46 ± 8 ms for the TrueBeam. Hence, the mean latencies determined by the two methods agreed within 13 ms for the ProBeam and within 2 ms for the TrueBeam.

CONCLUSIONS

A novel, simple and low-cost method for gating latency measurements that work across different radiotherapy platforms was demonstrated. Only the TrueBeam fully fulfilled the AAPM TG-142 recommendation of maximum 100 ms latencies.

Film measurement and analytical approach for assessing treatment accuracy and latency in a magnetic resonance-guided radiotherapy system

PURPOSE

We measure the dose distribution of gated delivery for different target motions and estimate the gating latency in a magnetic resonance-guided radiotherapy (MRgRT) system.

METHOD

The dose distribution accuracy of the gated MRgRT system (MRIdian, Viewray) was investigated using an in-house-developed phantom that was compatible with the magnetic field and gating method. This phantom contains a simulated tumor and a radiochromic film (EBT3, Ashland, Inc.). To investigate the effect of the number of beam switching and target velocity on the dose distribution, two types of target motions were applied. One is that the target was periodically moved at a constant velocity of 5 mm/s with different pause times (0, 1, 3, 10, and 20 s) between the motions. During different pause times, different numbers of beams were switched on/off. The other one is that the target was moved at velocities of 3, 5, 8, and 10 mm/s without any pause (i.e., continuous motion). The gated method was applied to these motions at MRIdian, and the dose distributions in each condition were measured using films. To investigate the relation between target motion and dose distribution in the gating method, we compared the results of the gamma analysis of the calculated and measured dose distributions. Moreover, we analytically estimated the gating latencies from the dose distributions measured using films and the gamma analysis results.

RESULTS

The gamma pass rate linearly decreased with increasing beam switching and target velocity. The overall gating latencies of beam-hold and beam-on were 0.51 ± 0.17 and 0.35 ± 0.05 s, respectively.

CONCLUSIONS

Film measurements highlighted the factors affecting the treatment accuracy of the gated MRgRT system. Our analytical approach, employing gamma analysis on films, can be used to estimate the overall latency of the gated MRgRT system.

Individual pulse monitoring and dose control system for pre-clinical implementation of FLASH-RT

Objective.Existing ultra-high dose rate (UHDR) electron sources lack dose rate independent dosimeters and a calibrated dose control system for accurate delivery. In this study, we aim to develop a custom single-pulse dose monitoring and a real-time dose-based control system for a FLASH enabled clinical linear accelerator (Linac).Approach.A commercially available point scintillator detector was coupled to a gated integrating amplifier and a real-time controller for dose monitoring and feedback control loop. The controller was programmed to integrate dose for each radiation pulse and stop the radiation beam when the prescribed dose was delivered. Additionally, the scintillator was mounted in a solid water phantom and placed underneath mice skin forin vivodose monitoring. The scintillator was characterized in terms of its radiation stability, mean dose-rate (Ḋm), and dose per pulse (Dp) dependence.Main results.TheDpexhibited a consistent ramp-up period across ∼4-5 pulse. The plastic scintillator was shown to be linear withḊm(40-380 Gy s-1) andDp(0.3-1.3 Gy Pulse-1) to within +/- 3%. However, the plastic scintillator was subject to significant radiation damage (16%/kGy) for the initial 1 kGy and would need to be calibrated frequently. Pulse-counting control was accurately implemented with one-to-one correspondence between the intended and the actual delivered pulses. The dose-based control was sufficient to gate on any pulse of the Linac.In vivodosimetry monitoring with a 1 cm circular cut-out revealed that during the ramp-up period, the averageDpwas ∼0.045 ± 0.004 Gy Pulse-1, whereas after the ramp-up it stabilized at 0.65 ± 0.01 Gy Pulse-1.Significance.The tools presented in this study can be used to determine the beam parameter space pertinent to the FLASH effect. Additionally, this study is the first instance of real-time dose-based control for a modified Linac at ultra-high dose rates, which provides insight into the tool required for future clinical translation of FLASH-RT.

Deformable abdominal phantom for the validation of real-time image guidance and deformable dose accumulation

PURPOSE

End-to-end testing with quality assurance (QA) phantoms for deformable dose accumulation and real-time image-guided radiotherapy (IGRT) has recently been recommended by American Association of Physicists in Medicine (AAPM) Task Groups 132 and 76. The goal of this work was to develop a deformable abdominal phantom containing a deformable three-dimensional dosimeter that could provide robust testing of these systems.

METHODS

The deformable abdominal phantom was fabricated from polyvinyl chloride plastisol and phantom motion was simulated with a programmable motion stage and plunger. A deformable normoxic polyacrylamide gel (nPAG) dosimeter was incorporated into the phantom apparatus to represent a liver tumor. Dosimeter data were acquired using magnetic resonance imaging (MRI). Static measurements were compared to planned dose distributions. Static and dynamic deformations were used to simulate inter- and intrafractional motion in the phantom and measurements were compared to baseline measurements.

RESULTS

The statically irradiated dosimeters matched the planned dose distribution with an average γ pass rates of 97.0 ± 0.5% and 97.5 ± 0.2% for 3%/5 mm and 5%/5 mm criteria, respectively. Static deformations caused measured dose distribution shifts toward the phantom plunger. During the dynamic deformation experiment, the dosimeter that utilized beam gating showed an improvement in the γ pass rate compared to the dosimeter that did not.

CONCLUSIONS

A deformable abdominal phantom apparatus which incorporates a deformable nPAG dosimeter was developed to test real-time IGRT systems and deformable dose accumulation algorithms. This apparatus was used to benchmark simple static irradiations in which it was found that measurements match well to the planned distributions. Deformable dose accumulation could be tested by directly measuring the shifts and blurring of the target dose due to interfractional organ deformation and motion. Dosimetric improvements were achieved from the motion management during intrafractional motion.

Markerless Pancreatic Tumor Target Localization Enabled By Deep Learning

PURPOSE

Deep learning is an emerging technique that allows us to capture imaging information beyond the visually recognizable level of a human being. Because of the anatomic characteristics and location, on-board target verification for radiation delivery to pancreatic tumors is a challenging task. Our goal was to use a deep neural network to localize the pancreatic tumor target on kV x-ray images acquired using an on-board imager for image guided radiation therapy.

METHODS AND MATERIALS

The network is set up in such a way that the input is either a digitally reconstructed radiograph image or a monoscopic x-ray projection image acquired by the on-board imager from a given direction, and the output is the location of the planning target volume in the projection image. To produce a sufficient number of training x-ray images reflecting the vast number of possible clinical scenarios of anatomy distribution, a series of changes were introduced to the planning computed tomography images, including deformation, rotation, and translation, to simulate inter- and intrafractional variations. After model training, the accuracy of the model was evaluated by retrospectively studying patients who underwent pancreatic cancer radiation therapy. Statistical analysis using mean absolute differences (MADs) and Lin's concordance correlation coefficient were used to assess the accuracy of the predicted target positions.

RESULTS

MADs between the model-predicted and the actual positions were found to be less than 2.60 mm in anteroposterior, lateral, and oblique directions for both axes in the detector plane. For comparison studies with and without fiducials, MADs are less than 2.49 mm. For all cases, Lin's concordance correlation coefficients between the predicted and actual positions were found to be better than 93%, demonstrating the success of the proposed deep learning for image guided radiation therapy.

CONCLUSIONS

We demonstrated that markerless pancreatic tumor target localization is achievable with high accuracy by using a deep learning technique approach.

Towards real-time respiratory motion prediction based on long short-term memory neural networks

Radiation therapy of thoracic and abdominal tumors requires incorporating the respiratory motion into treatments. To precisely account for the patient's respiratory motions and predict the respiratory signals, a generalized model for predictions of different types of patients' respiratory motions is desired. The aim of this study is to explore the feasibility of developing a long short-term memory (LSTM)-based generalized model for the respiratory signal prediction. To achieve that, 1703 sets of real-time position management (RPM) data were collected from retrospective studies across three clinical institutions. These datasets were separated as the training, internal validity and external validity groups. Among all the datasets, 1187 datasets were used for model development and the remaining 516 datasets were used to test the model's generality power. Furthermore, an exhaustive grid search was implemented to find the optimal hyper-parameters of the LSTM model. The hyper-parameters are the number of LSTM layers, the number of hidden units, the optimizer, the learning rate, the number of epochs, and the length of time lags. The obtained model achieved superior accuracy over conventional artificial neural network (ANN) models: with the prediction window equaling to 500 ms, the LSTM model achieved an average relative mean absolute error (MAE) of 0.037, an average root mean square error (RMSE) of 0.048, and a maximum error (ME) of 1.687 in the internal validity data, and an average relative MAE of 0.112, an average RMSE of 0.139 and an ME of 1.811 in the external validity data. Compared to the LSTM model trained with default hyper-parameters, the MAE of the optimized model results decreased by 20%, indicating the importance of tuning the hyper-parameters of LSTM models to obtain superior accuracies. This study demonstrates the potential of deep LSTM models for the respiratory signal prediction and illustrates the impacts of major hyper-parameters in LSTM models.