Prediction of the position of external markers using a recurrent neural network trained with unbiased online recurrent optimization for safe lung cancer radiotherapy.

BACKGROUND AND OBJECTIVE: During lung cancer radiotherapy, the position of infrared reflective objects on the chest can be recorded to estimate the tumor location. However, radiotherapy systems have a latency inherent to robot control limitations that impedes the radiation delivery precision. Prediction with online learning of recurrent neural networks (RNN) allows for adaptation to non-stationary respiratory signals, but classical methods such as real-time recurrent learning (RTRL) and truncated backpropagation through time are respectively slow and biased. This study investigates the capabilities of unbiased online recurrent optimization (UORO) to forecast respiratory motion and enhance safety in lung radiotherapy. METHODS: We used nine observation records of the three-dimensional (3D) position of three external markers on the chest and abdomen of healthy individuals breathing during intervals from 73s to 222s. The sampling frequency was 10Hz, and the amplitudes of the recorded trajectories range from 6mm to 40mm in the superior-inferior direction. We forecast the 3D location of each marker simultaneously with a horizon value (the time interval in advance for which the prediction is made) between 0.1s and 2.0s, using an RNN trained with UORO. We compare its performance with an RNN trained with RTRL, least mean squares (LMS), and offline linear regression. We provide closed-form expressions for quantities involved in the gradient loss calculation in UORO, thereby making its implementation efficient. Training and cross-validation were performed during the first minute of each sequence. RESULTS: On average over the horizon values considered and the nine sequences, UORO achieves the lowest root-mean-square (RMS) error and maximum error among the compared algorithms. These errors are respectively equal to 1.3mm and 8.8mm, and the prediction time per time step was lower than 2.8ms (Dell Intel core i9-9900K 3.60 GHz). Linear regression has the lowest RMS error for the horizon values 0.1s and 0.2s, followed by LMS for horizon values between 0.3s and 0.5s, and UORO for horizon values greater than 0.6s. CONCLUSIONS: UORO can accurately predict the 3D position of external markers for intermediate to high response times with an acceptable time performance. This will help limit unwanted damage to healthy tissues caused by radiotherapy.

A Real-Time Four-Dimensional Reconstruction Algorithm of Cine-Magnetic Resonance Imaging (Cine-MRI) Using Deep Learning.

Purpose The purpose of this study is to propose algorithms and methods for achieving high accuracy in tracking and interception irradiation technology for tumors that move by respiration using MR-linac (MRIdian®, ViewRay Inc.) and to use deep learning to predict the movement of moving tumors in real time during radiation therapy and reconstruct cine magnetic resonance imaging (cine-MRI) into four-dimensional (4D) movies. Methods In this study, we propose a reconstruction algorithm using 4DCT for treatment planning taken before irradiation as training data in consideration of the actual treatment flow. In the algorithm, two neural networks made before treatment are used to reconstruct 4D movies that predict tumor movement in real time during treatment. Cycle GAN (generative adversarial network) was used to convert MR images to CT images, and long short-term memory was used to convert cine-MRI to 4D movies and predict tumor movement. Results We succeeded in predicting the time including the imaging time of the MR images, the lag until irradiation, and the calculation time in the algorithm. In addition, the conversion and prediction results at each phase of reconstruction were generally good so that they could be clinically applied. Conclusions The reconstruction algorithm proposed in this study enables high-precision radiotherapy while predicting the volume information of the tumor and the actual tumor position, which could not be obtained during radiotherapy.

Prediction of the motion of chest internal points using a recurrent neural network trained with real-time recurrent learning for latency compensation in lung cancer radiotherapy.

During the radiotherapy treatment of patients with lung cancer, the radiation delivered to healthy tissue around the tumor needs to be minimized, which is difficult because of respiratory motion and the latency of linear accelerator (LINAC) systems. In the proposed study, we first use the Lucas-Kanade pyramidal optical flow algorithm to perform deformable image registration (DIR) of chest computed tomography (CT) scan images of four patients with lung cancer. We then track three internal points close to the lung tumor based on the previously computed deformation field and predict their position with a recurrent neural network (RNN) trained using real-time recurrent learning (RTRL) and gradient clipping. The breathing data is quite regular, sampled at approximately 2.5 Hz, and includes artificially added drift in the spine direction. The amplitude of the motion of the tracked points ranged from 12.0 mm to 22.7 mm. Finally, we propose a simple method for recovering and predicting three-dimensional (3D) tumor images from the tracked points and the initial tumor image, based on a linear correspondence model and the Nadaraya-Watson non-linear regression. The root-mean-square (RMS) error, maximum error and jitter corresponding to the RNN prediction on the test set were smaller than the same performance measures obtained with linear prediction and least mean squares (LMS). In particular, the maximum prediction error associated with the RNN, equal to 1.51 mm, is respectively 16.1% and 5.0% lower than the error given by a linear predictor and LMS. The average prediction time per time step with RTRL is equal to 119 ms, which is less than the 400 ms marker position sampling time. The tumor position in the predicted images appears visually correct, which is confirmed by the high mean cross-correlation between the original and predicted images, equal to 0.955. The standard deviation of the Gaussian kernel and the number of layers in the optical flow algorithm were the parameters having the most significant impact on registration performance. Their optimization led respectively to a 31.3% and 36.2% decrease in the registration error. Using only a single layer proved to be detrimental to the registration quality because tissue motion in the lower part of the lung has a high amplitude relative to the resolution of the CT scan images. The random initialization of the hidden units and the number of these hidden units were found to be the most important factors affecting the performance of the RNN. Increasing the number of hidden units from 15 to 250 led to a 56.3% decrease in the prediction error on the cross-validation data. Similarly, optimizing the standard deviation of the initial Gaussian distribution of the synaptic weights σinitRNN led to a 28.4% decrease in the prediction error on the cross-validation data, with the error minimized for σinitRNN=0.02 with the four patients.

Development of a markerless tumor-tracking algorithm using prior four-dimensional cone-beam computed tomography.

Respiratory motion management is a huge challenge in radiation therapy. Respiratory motion induces temporal anatomic changes that distort the tumor volume and its position. In this study, a markerless tumor-tracking algorithm was investigated by performing phase recognition during stereotactic body radiation therapy (SBRT) using four-dimensional cone-beam computer tomography (4D-CBCT) obtained at patient registration, and in-treatment cone-beam projection images. The data for 20 treatment sessions (five lung cancer patients) were selected for this study. Three of the patients were treated with conventional flattening filter (FF) beams, and the other two were treated with flattening filter-free (FFF) beams. Prior to treatment, 4D-CBCT was acquired to create the template projection images for 10 phases. In-treatment images were obtained at near real time during treatment. Template-based phase recognition was performed for 4D-CBCT re-projected templates using prior 4D-CBCT based phase recognition algorithm and was compared with the results generated by the Amsterdam Shroud (AS) technique. Visual verification technique was used for the verification of the phase recognition and AS technique at certain tumor-visible angles. Offline template matching analysis using the cross-correlation indicated that phase recognition performed using the prior 4D-CBCT and visual verification matched up to 97.5% in the case of FFF, and 95% in the case of FF, whereas the AS technique matched 83.5% with visual verification for FFF and 93% for FF. Markerless tumor tracking based on phase recognition using prior 4D-CBCT has been developed successfully. This is the first study that reports on the use of prior 4D-CBCT based on normalized cross-correlation technique for phase recognition.

Quantification of susceptibility change at high-concentrated SPIO-labeled target by characteristic phase gradient recognition.

Phase map cross-correlation detection and quantification may produce highlighted signal at superparamagnetic iron oxide nanoparticles, and distinguish them from other hypointensities. The method may quantify susceptibility change by performing least squares analysis between a theoretically generated magnetic field template and an experimentally scanned phase image. Because characteristic phase recognition requires the removal of phase wrap and phase background, additional steps of phase unwrapping and filtering may increase the chance of computing error and enlarge the inconsistence among algorithms. To solve problem, phase gradient cross-correlation and quantification method is developed by recognizing characteristic phase gradient pattern instead of phase image because phase gradient operation inherently includes unwrapping and filtering functions. However, few studies have mentioned the detectable limit of currently used phase gradient calculation algorithms. The limit may lead to an underestimation of large magnetic susceptibility change caused by high-concentrated iron accumulation. In this study, mathematical derivation points out the value of maximum detectable phase gradient calculated by differential chain algorithm in both spatial and Fourier domain. To break through the limit, a modified quantification method is proposed by using unwrapped forward differentiation for phase gradient generation. The method enlarges the detectable range of phase gradient measurement and avoids the underestimation of magnetic susceptibility. Simulation and phantom experiments were used to quantitatively compare different methods. In vivo application performs MRI scanning on nude mice implanted by iron-labeled human cancer cells. Results validate the limit of detectable phase gradient and the consequent susceptibility underestimation. Results also demonstrate the advantage of unwrapped forward differentiation compared with differential chain algorithms for susceptibility quantification at high-concentrated iron accumulation.

Tumor growth suppression by gadolinium-neutron capture therapy using gadolinium-entrapped liposome as gadolinium delivery agent.

Neutron capture therapy (NCT) is a promising non-invasive cancer therapy approach and some recent NCT research has focused on using compounds containing gadolinium as an alternative to currently used boron-10 considering several advantages that gadolinium offers compared to those of boron. In this study, we evaluated gadolinium-entrapped liposome compound as neutron capture therapy agent by in vivo experiment on colon-26 tumor-bearing mice. Gadolinium compound were injected intravenously via tail vein and allowed to accumulate into tumor site. Tumor samples were taken for quantitative analysis by ICP-MS at 2, 12, and 24 h after gadolinium compound injection. Highest gadolinium concentration was observed at about 2 h after gadolinium compound injection with an average of 40.3 µg/g of wet tumor tissue. We performed neutron irradiation at JRR-4 reactor facility of Japan Atomic Energy Research Institute in Tokaimura with average neutron fluence of 2×10¹² n/cm². The experimental results showed that the tumor growth suppression of gadolinium-injected irradiated group was revealed until about four times higher compared to the control group, and no significant weight loss were observed after treatment suggesting low systemic toxicity of this compound. The gadolinium-entrapped liposome will become one of the candidates for Gd delivery system on NCT.

A study of internal defect testing with the laser-EMAT ultrasonic method.

This paper studies the ultrasonic detection and evaluation of internal volume defects in metals using laser generation and electromagnetic acoustic transducer (EMAT) detection. A finite element model is developed to simulate the interaction of laser-generated ultrasonic waves with the defect in the material. Not only have the directly scattered shear waves been observed, but also the mode-converted creeping waves on the defect surface. A noncontact laser-EMAT ultrasonic testing experimental system was successfully applied to validate the observed phenomena in the simulation results. The defect can not only be detected and located by the directly scattered shear waves, but can also be quickly evaluated with a new method based on quantitative time-of-flight analysis of the directly scattered waves and the mode-converted waves on the defect surface.

Correction of chemical shift misregistration by images from two different bandwidths.

One major effect caused by the different chemical shift frequencies of water and fat is the misregistration between the two components in MR images. Methods to correct misregistration are required in clinical MRI for accurate localization and artifact reduction. One of the methods uses the images scanned at opposite readout gradients to separate water and fat signal in the k-space. Its signal-to-noise ratio (SNR) achieves maximum when misregistration is around 0.9 pixels and deteriorates rapidly as the misregistration gets larger. In this work, we proposed a method to correct the chemical shift misregistration by using two data sets acquired at two different bandwidths. It is more generalized and flexible than the former method of opposite readout gradients and covers the former one as a special case. In both simulation and experiment, the new method is proved to be capable of correcting large chemical shift misregistration and maintain a good SNR.

Phase gradient imaging for positive contrast generation to superparamagnetic iron oxide nanoparticle-labeled targets in magnetic resonance imaging.

Positive contrast imaging methods produce enhanced signal at large magnetic field gradient in magnetic resonance imaging. Several postprocessing algorithms, such as susceptibility gradient mapping and phase gradient mapping methods, have been applied for positive contrast generation to detect the cells targeted by superparamagnetic iron oxide nanoparticles. In the phase gradient mapping methods, smoothness condition has to be satisfied to keep the phase gradient unwrapped. Moreover, there has been no discussion about the truncation artifact associated with the algorithm of differentiation that is performed in k-space by the multiplication with frequency value. In this work, phase gradient methods are discussed by considering the wrapping problem when the smoothness condition is not satisfied. A region-growing unwrapping algorithm is used in the phase gradient image to solve the problem. In order to reduce the truncation artifact, a cosine function is multiplied in the k-space to eliminate the abrupt change at the boundaries. Simulation, phantom and in vivo experimental results demonstrate that the modified phase gradient mapping methods may produce improved positive contrast effects by reducing truncation or wrapping artifacts.