Personalized Respiratory Motion Model Using Conditional Generative Networks for MR-Guided Radiotherapy

Respiratory motion during PET imaging causes the resulting images to become corrupted by artifacts. Respiratory motion models have found application in estimating and correcting for the effects of motion in a wide range of applications. In this work we evaluate the performance of an MR-derived motion model for motion-correction of PET data that is suitable for use in a simultaneous PET-MR imaging system. The technique is based on the formation of a subject-specific respiratory motion model from near real-time dynamic MR images and is capable of making real-time motion estimates based on an MR navigator signal. This approach has an advantage over gating based approaches in that individual motion corrections can be applied at each time point, avoiding the need to combine large amounts of PET data for correction by a single transformation for each gate. In our investigation, we evaluate the effects of the MR-derived motion model and PET motion compensation using dynamic PET simulations based on dynamic MR data (105 real-time frames, 0.7s each) of the thorax acquired from a healthy volunteer. This large amount of data, acquired during different breathing patterns, allows us to more accurately simulate the complex respiratory motion variation that is observed in real PET acquisitions. PET images are corrected for motion using two approaches: reconstruct-transform-average (RTA) and motion compensated image reconstruction (MCIR). We compare the methods in terms of quantification and visualization of three tumors that were artificially added to the emission and attenuation maps of the volunteer close to the diaphragm. It is demonstrated that our motion model can be used to correct successfully for motion in real-time dynamic PET-MR acquisitions. The correction is feasible using either RTA or MCIR, although the latter provides

Emission guided radiation therapy (EGRT) is a new modality that uses PET emissions in real-time for direct tumor tracking during radiation delivery. Radiation beamlets are delivered along positron emission tomography (PET) lines of response (LORs) by a fast rotating ring therapy unit consisting of a linear accelerator (Linac) and PET detectors. The feasibility of tumor tracking and a primitive modulation method to compensate for attenuation have been demonstrated using a 4D digital phantom in our prior work. However, the essential capability of achieving dose modulation as in conventional intensity modulated radiation therapy (IMRT) treatments remains absent. In this work, the authors develop a planning scheme for EGRT to accomplish sophisticated intensity modulation based on an IMRT plan while preserving tumor tracking. The planning scheme utilizes a precomputed LOR response probability distribution to achieve desired IMRT planning modulation with effects of inhomogeneous attenuation and nonuniform background activity distribution accounted for. Evaluation studies are performed on a 4D digital patient with a simulated lung tumor and a clinical patient who has a moving breast cancer metastasis in the lung. The Linac dose delivery is simulated using a voxel-based Monte Carlo algorithm. The IMRT plan is optimized for a planning target volume (PTV) that encompasses the tumor motion using the MOSEK package and a Pinnacle3™ workstation (Philips Healthcare, Fitchburg, WI) for digital and clinical patients, respectively. To obtain the emission data for both patients, the Geant4 application for tomographic emission (GATE) package and a commercial PET scanner are used. As a comparison, 3D and helical IMRT treatments covering the same PTV based on the same IMRT plan are simulated. 3D and helical IMRT treatments show similar dose distribution. In the digital patient case, compared with the 3D IMRT treatment, EGRT achieves a 15.1% relative increase in dose to 95% of the gross tumor volume (GTV) and a 31.8% increase to 50% of the GTV. In the patient case, EGRT yields a 15.2% relative increase in dose to 95% of the GTV and a 20.7% increase to 50% of the GTV. The organs at risk (OARs) doses are kept similar or lower for EGRT in both cases. Tumor tracking is observed in the presence of planning modulation in all EGRT treatments. As compared to conventional IMRT treatments, the proposed EGRT planning scheme allows an escalated target dose while keeping dose to the OARs within the same planning limits. With the capabilities of incorporating planning modulation and accurate tumor tracking, EGRT has the potential to greatly improve targeting in radiation therapy and enable a practical and effective implementation of 4D radiation therapy for planning and delivery.

Purpose:Dynamic tumor tracking or motion compensation techniques have proposed to modify beam delivery following lung tumor motion on the flight. Conventional treatment plan QA could be performed in advance since every delivery may be different. Markerless lung tumor tracking using beams eye view EPID images provides a best treatment evaluation mechanism. The purpose of this study is to improve the accuracy of the online markerless lung tumor motion tracking method.Methods:The lung tumor could be located on every frame of MV images during radiation therapy treatment by comparing with corresponding digitally reconstructed radiograph (DRR). A kV‐MV CT corresponding curve is applied on planning kV CT to generate MV CT images for patients in order to enhance the similarity between DRRs and MV treatment images. This kV‐MV CT corresponding curve was obtained by scanning a same CT electron density phantom by a kV CT scanner and MV scanner (Tomotherapy) or MV CBCT. Two sets of MV DRRs were then generated for tumor and anatomy without tumor as the references to tracking the tumor on beams eye view EPID images.Results:Phantom studies were performed on a Varian TrueBeam linac. MV treatment images were acquired continuously during each treatment beam delivery at 12 gantry angles by iTools. Markerless tumor tracking was applied with DRRs generated from simulated MVCT. Tumors were tracked on every frame of images and compared with expected positions based on programed phantom motion. It was found that the average tracking error were 2.3 mm.Conclusion:This algorithm is capable of detecting lung tumors at complicated environment without implanting markers. It should be noted that the CT data has a slice thickness of 3 mm. This shows the statistical accuracy is better than the spatial accuracy.This project has been supported by a Varian Research Grant.

Purpose:To detect target position on kV X‐ray fluoroscopic images using a feature‐based tracking algorithm, Accelerated‐KAZE (AKAZE), for markerless real‐time tumor tracking (RTTT).Methods:Twelve lung cancer patients treated with RTTT on the Vero4DRT (Mitsubishi Heavy Industries, Japan, and Brainlab AG, Feldkirchen, Germany) were enrolled in this study. Respiratory tumor movement was greater than 10 mm. Three to five fiducial markers were implanted around the lung tumor transbronchially for each patient. Before beam delivery, external infrared (IR) markers and the fiducial markers were monitored for 20 to 40 s with the IR camera every 16.7 ms and with an orthogonal kV x‐ray imaging subsystem every 80 or 160 ms, respectively. Target positions derived from the fiducial markers were determined on the orthogonal kV x‐ray images, which were used as the ground truth in this study. Meanwhile, tracking positions were identified by AKAZE. Among a lot of feature points, AKAZE found high‐quality feature points through sequential cross‐check and distance‐check between two consecutive images. Then, these 2D positional data were converted to the 3D positional data by a transformation matrix with a predefined calibration parameter. Root mean square error (RMSE) was calculated to evaluate the difference between 3D tracking and target positions. A total of 393 frames was analyzed. The experiment was conducted on a personal computer with 16 GB RAM, Intel Core i7‐2600, 3.4 GHz processor.Results:Reproducibility of the target position during the same respiratory phase was 0.6 +/− 0.6 mm (range, 0.1–3.3 mm). Mean +/− SD of the RMSEs was 0.3 +/− 0.2 mm (range, 0.0–1.0 mm). Median computation time per frame was 179 msec (range, 154–247 msec).Conclusion:AKAZE successfully and quickly detected the target position on kV X‐ray fluoroscopic images. Initial results indicate that the differences between 3D tracking and target position would be clinically acceptable.

Purpose: To evaluate feasibility of markerless 3D tumor trajectory tracking in CBCT projections using a digital subtraction technique for tumor enhancement in low contrast lesions. Method and Materials: 3D tumor position was estimated in CBCT measured projections (MP) by registering digital subtraction projections with templates. A fan beam 4DCT scan served as the planning CT (pRCCT), from which two sets of DRRs were generated using an in‐house GPU‐based algorithm. Tumor templates (TT) were generated by masking the GTV+0.5cm and projecting at the MP view angles for each 4D phase image. Subtraction templates (ST) were generated by first removing the GTV+0.5cm from the pRCCT, positioning the pRCCT at the pose of the patient during CBCT acquisition, and projecting at the MP angles for each 4D phase image. The ST was subtracted from the corresponding MP to create subtraction projections (SP). The 2D tumor position in each SP was calculated by matching with the corresponding TT using a robust correlation metric, and transformed to 3D positions by backprojection with a prior depth estimate. 3D position error was assessed by comparing to 3D positions of implanted markers, estimated with a separate algorithm, in a clinical scan. Results: The mean absolute error for the conventional method was 2.1mm, 2.5mm, and 2.9mm in mediolateral, anterior‐posterior, and superior‐inferior. The 90% error level was 4.1mm, 6.4mm, and 4.5mm. For the digital subtraction method, the mean errors were 0.8mm, 0.5mm, and 0.6mm and 90% errors were 1.4mm, 0.9mm, and 1mm. Conclusions: Digital subtraction is feasible for tumor enhancement to reduce error in tumor trajectory measurement. A framework for validation of markerless tracking algorithms using implanted markers was developed. Supported by NIH R01CA116249.

Purpose: This study aimed to evaluate the potential and feasibility of positron emission tomography (PET) for dynamic tumor tracking during radiation treatment. We have proposed a center of mass (CoM) tumor tracking algorithm and investigated the geometric accuracy of the algorithm. Methods: The proposed PET dynamic tumor tracking algorithm estimated the target position information through the CoM of the segmented target volume on gated PET images which were continuously updated throughout a scan. External respiratory motion and list‐mode PET data were acquired from a phantom programmed to move with measured respiratory traces. The phantom was cylindrical with six sphere targets (10, 13, 17, 22, 28, and 37 mm in diameter). To test the algorithm on respiratory motion with clinical variability, respiratory traces of higher than average magnitudes and complexity were selected from a large database. The traces were 3D‐measured motion representing typical motion (T1), baseline variations (T2), predominantly left‐right motion (T3), and high‐frequency breathing (T4). Estimation errors were quantified through Euclidean distances between the motion traces and estimated CoM trajectories. Results: The overall time‐averaged error was 1.6 mm over all traces (T1‐T4) of all targets. There were small variations in the errors between the traces, although they were very distinct. The overall time‐averaged error of each trace (T1‐T4) was 1.3, 1.3, 1.7, and 2.2 mm, respectively. The accuracy of the estimates was consistent for all targets except the smallest one: For the 17 mm target, the overall time‐averaged errors of T1‐T4 were 1.2, 1.2, 1.4, and 1.7 mm, respectively; For the 10 mm target, those of T1‐T4 were 2.3, 2.2, 2.6, and 3.9 mm, respectively. Conclusion: We developed an algorithm using list‐mode PET data and external respiratory motion data for dynamic tumor tracking in a combined PET‐linac system. The overall tracking error in phantom studies was less than 2 mm. Research supported by the Kwanjeong Educational Foundation, NIH/NCI R01 93626, Stanford Bio‐X, NHMRC Australia Fellowship, NIH/NCI through SBIR with RefleXion Medical R43CA153466.

Accurate state estimation for quadrotors under wind-induced disturbances remains a critical challenge in dynamic outdoor environments. Existing model-based and data-driven approaches often struggle with real-time adaptation and catastrophic forgetting when faced with continuous wind disturbances. This paper proposes an online continual physics-informed learning framework that integrates physics-informed neural networks with continual backpropagation to address these limitations. The physics-informed neural networks architecture embeds quadrotor dynamics into the neural network training process, ensuring physical consistency, while continual backpropagation enables continual learning from real-time streaming data without compromising previously acquired knowledge. Experimental validation on a simulation platform demonstrates the accuracy and robustness of the framework in ideal and wind-disturbed scenarios.

BackgroundCine cardiovascular magnetic resonance (CMR) is challenging in patients who cannot perform repeated breath holds. Real-time, free-breathing acquisition is an alternative, but image quality is typically inferior. There is a clinical need for techniques that achieve similar image quality to the segmented cine using a free breathing acquisition. Previously, high quality retrospectively gated cine images have been reconstructed from real-time acquisitions using parallel imaging and motion correction. These methods had limited clinical applicability due to lengthy acquisitions and volumetric measurements obtained with such methods have not previously been evaluated systematically.MethodsThis study introduces a new retrospective reconstruction scheme for real-time cine imaging which aims to shorten the required acquisition. A real-time acquisition of 16-20s per acquired slice was inputted into a retrospective cine reconstruction algorithm, which employed non-rigid registration to remove respiratory motion and SPIRiT non-linear reconstruction with temporal regularization to fill in missing data. The algorithm was used to reconstruct cine loops with high spatial (1.3-1.8 × 1.8-2.1 mm2) and temporal resolution (retrospectively gated, 30 cardiac phases, temporal resolution 34.3 ± 9.1 ms). Validation was performed in 15 healthy volunteers using two different acquisition resolutions (256 × 144/192 × 128 matrix sizes). For each subject, 9 to 12 short axis and 3 long axis slices were imaged with both segmented and real-time acquisitions. The retrospectively reconstructed real-time cine images were compared to a traditional segmented breath-held acquisition in terms of image quality scores. Image quality scoring was performed by two experts using a scale between 1 and 5 (poor to good). For every subject, LAX and three SAX slices were selected and reviewed in the random order. The reviewers were blinded to the reconstruction approach and acquisition protocols and scores were given to segmented and retrospective cine series. Volumetric measurements of cardiac function were also compared by manually tracing the myocardium for segmented and retrospective cines.ResultsMean image quality scores were similar for short axis and long axis views for both tested resolutions. Short axis scores were 4.52/4.31 (high/low matrix sizes) for breath-hold vs. 4.54/4.56 for real-time (paired t-test, P = 0.756/0.011). Long axis scores were 4.09/4.37 vs. 3.99/4.29 (P = 0.475/0.463). Mean ejection fraction was 60.8/61.4 for breath-held acquisitions vs. 60.3/60.3 for real-time acquisitions (P = 0.439/0.093). No significant differences were seen in end-diastolic volume (P = 0.460/0.268) but there was a trend towards a small overestimation of end-systolic volume of 2.0/2.5 ml, which did not reach statistical significance (P = 0.052/0.083).ConclusionsReal-time free breathing CMR can be used to obtain high quality retrospectively gated cine images in 16-20s per slice. Volumetric measurements and image quality scores were similar in images from breath-held segmented and free breathing, real-time acquisitions. Further speedup of image reconstruction is still needed.

RT-SRTS: Angle-agnostic real-time simultaneous 3D reconstruction and tumor segmentation from single X-ray projection

The aim of the study is to optimize an existing microwave powered remote plasma source (RPS) with respect to the etching rate and gas temperature and to simplify the setup to save production costs. The RPS, which is shown in figure 1, is a low-pressure plasma source where the plasma is generated and exists mainly in the chamber of the source. Only radicals migrate out of the RPS. This is one important feature, that the plasma source is used for etching processes when ion bombardment and high thermal strain of the substrate must be prevented. The etching process is a chemical process, where the radicals react with the substrate surface atoms forming gaseous molecules. The benefit is a damage-free, dry and clean substrate surface. To achieve these goals, a FEM-based model of the RPS has been developed to investigate the microwave distribution and the microwave coupling into the plasma chamber, as well as the plasma itself. In this paper different examples of FEM based microwave simulations by different conditions and their experimental validations will be presented. To compare the calculated electric field distribution in the RPS with the real field distribution, PMMA-substrates were placed inside the plasma chamber of the source. They are heated up by the electric field and evaluated with an infrared camera and liquid crystal sheets. Both the measured and the calculated field distribution show a very good conformity. When the electric field is high enough in the plasma chamber the plasma ignites, the electron density and thus the permittivity and the conductivity increase, which changes again the electric field distribution. For this purpose, the FEM-model has been extended by the Drude model1. The model considers the equation of motion with a damping term for the electrons, leading to an expression for the conductivity. Results for various electron densities as well as their corresponding electric field distributions are presented and compared with optical measurements. Fig. 1. The figure shows the scheme of the RPS with its main components and functions.

IntroductionDynamic tumor tracking (DTT) is a motion management technique where the radiation beam follows a moving tumor in real time. Not modelling DTT beam motion in the treatment planning system leaves an organ at risk (OAR) vulnerable to exceeding its dose limit. This work investigates two planning strategies for DTT plans, the “Boolean OAR Method” and the “Aperture Sorting Method,” to determine if they can successfully spare an OAR while maintaining sufficient target coverage.Materials and methodsA step‐and‐shoot intensity modulated radiation therapy (sIMRT) treatment plan was re‐optimized for 10 previously treated liver stereotactic ablative radiotherapy patients who each had one OAR very close to the target. Two planning strategies were investigated to determine which is more effective at sparing an OAR while maintaining target coverage: (1) the “Boolean OAR Method” created a union of an OAR's contours from two breathing phases (exhale and inhale) on the exhale phase (the planning CT) and protected this combined OAR during plan optimization, (2) the “Aperture Sorting Method” assigned apertures to the breathing phase where they contributed the least to an OAR's maximum dose.ResultsAll 10 OARs exceeded their dose constraints on the original plan four‐dimensional (4D) dose distributions and average target coverage was V100% = 91.3% ± 2.9% (ranging from 85.1% to 94.8%). The “Boolean OAR Method” spared 7/10 OARs, and mean target coverage decreased to V100% = 87.1% ± 3.8% (ranging from 80.7% to 93.7%). The “Aperture Sorting Method” spared 9/10 OARs and the mean target coverage remained high at V100% = 91.7% ± 2.8% (ranging from 84.9% to 94.5%).Conclusions4D planning strategies are simple to implement and can improve OAR sparing during DTT treatments. The “Boolean OAR Method” improved sparing of OARs but target coverage was reduced. The “Aperture Sorting Method” further improved sparing of OARs and maintained target coverage.

Patients with locally advanced prostate cancer require the prostate and pelvic lymph nodes to be irradiated simultaneously during radiation therapy treatment. However, relative motion between treatment targets decreases dosimetric conformity. Current treatment methods mitigate this error by having large treatment margins and often prioritize the prostate at patient setup at the cost of lymph node coverage. Treatment accuracy can be improved through real-time multi-target adaptation which requires simultaneous motion monitoring of both the prostate and lymph node targets. This study developed and evaluated an intrafraction pelvic bone motion monitoring method as a surrogate for pelvic lymph node displacement to be combined with prostate motion monitoring to enable multi-target six-degrees-of-freedom (6DoF) tracking using 2D kV projections acquired during treatment. A method to monitor pelvic bone translation and rotation was developed and retrospectively applied to images from 20 patients treated in the TROG 15.01 Stereotactic Prostate Ablative Radiotherapy with Kilovoltage Intrafraction Monitoring (KIM) trial. The pelvic motion monitoring method performed template matching to calculate the 6DoF position of the pelvis from 2D kV images. The method first generated a library of digitally reconstructed radiographs (DRRs) for a range of imaging angles and pelvic rotations. The normalized 2D cross-correlations were then calculated for each incoming kV image and a subset of DRRs and the DRR with the maximum correlation coefficient was used to estimate the pelvis translation and rotation. Translation of the pelvis in the unresolved direction was calculated using a 3D Gaussian probability estimation method. Prostate motion was measured using the KIM marker tracking method. The pelvic motion monitoring method was compared to the ground truth obtained from a 6DoF rigid registration of the CBCT and CT. The geometric errors of the pelvic motion monitoring method demonstrated sub-mm and sub-degree accuracy and precision in the translational directions ( , , ) and rotational directions ( , , ). The 3D relative displacement between the prostate and pelvic bones exceeded 2, 3, 5, and 7mm for approximately 66%, 44%, 12%, and 7% of the images. Accurate intrafraction pelvic bone motion monitoring in 6DoF was demonstrated on 2D kV images, providing a necessary tool for real-time multi-target motion-adapted treatment.

In this paper we propose an integrated cardiac segmentation and motion tracking algorithm. First, we present a subject-specific dynamical model (SSDM) that simultaneously handles inter-subject variability and temporal dynamics (intra-subject variability), such that it can progressively identify the subject vector associated with a new cardiac sequence, and use this subject vector to predict the subject-specific segmentation of the future frames based on the shapes observed in earlier frames. Second, we use the segmentation as a guide in selecting feature points with significant shape characteristics, and invoke the generalized robust point matching (G-RPM) strategy with boundary element method (BEM)-based regularization model to estimate physically realistic displacement field in a computationally efficient way. The integrated algorithm is formulated in a recursive Bayesian framework that sequentially segments cardiac images and estimates myocardial displacements. ldquoLeave-one-outrdquo validation on 32 sequences demonstrates that the segmentation results are improved when the SSDM is used, and the tracking results are much more accurate when the segmentation module is added.

Robotic devices have great potential in physical therapy owing to their repeatability, reliability and cost economy. However, there are great challenges to realize active control strategy, since the operator’s motion intention is uneasy to be recognized by robotics online. The purpose of this paper is to propose a subject-specific electromyography (EMG)-driven musculoskeletal model to estimate subject’s joint torque in real time, which can be used to detect his/her motion intention by forward dynamics, and then to explore its potential applications in rehabilitation robotics control. The musculoskeletal model uses muscle activation dynamics to extract muscle activation from raw EMG signals, a Hill-type muscle-tendon model to calculate muscle contraction force, and a proposed subject-specific musculoskeletal geometry model to calculate muscular moment arm. The parameters of muscle activation dynamics and muscle-tendon model are identified by off-line optimization methods in order to minimize the differences between the estimated muscular torques and the reference torques. Validation experiments were conducted on six healthy subjects to evaluate the proposed model. Experimental results demonstrated the model’s ability to predict knee joint torque with the coefficient of determination ([Formula: see text] value of [Formula: see text] and the normalized root-mean-square error (RMSE) of [Formula: see text].

Dynamic hysteresis loop shapes and magnetic power losses are studied in nonoriented Fe-Si laminations exhibiting significant excess losses. Measurements are carried out under controlled sinusoidal induction in the frequency range from 1 Hz to 1.6 kHz, at various peak inductions from 0.25 to 1.5 T. Excess losses are found to obey a f3/2 law up to frequencies of 200–400 Hz, depending on peak induction. Beyond this limit, definite deviations are observed, due to eddy current shielding. Detailed information on the flux and field distribution in this high frequency regime is obtained by finite element solutions of Maxwell equations employing the dynamic Preisach model to describe quasi-static hysteresis and dynamic wall processes. The agreement between theoretical predictions and measurements is discussed.