4D deep learning for real-time volumetric optical coherence elastography.

PURPOSE: Elasticity of soft tissue provides valuable information to physicians during treatment and diagnosis of diseases. A number of approaches have been proposed to estimate tissue stiffness from the shear wave velocity. Optical coherence elastography offers a particularly high spatial and temporal resolution. However, current approaches typically acquire data at different positions sequentially, making it slow and less practical for clinical application. METHODS: We propose a new approach for elastography estimations using a fast imaging device to acquire small image volumes at rates of 831 Hz. The resulting sequence of phase image volumes is fed into a 4D convolutional neural network which handles both spatial and temporal data processing. We evaluate the approach on a set of image data acquired for gelatin phantoms of known elasticity. RESULTS: Using the neural network, the gelatin concentration of unseen samples was predicted with a mean error of 0.65 ± 0.81 percentage points from 90 subsequent volumes of phase data only. We achieve a data acquisition and data processing time of under 12 ms and 22 ms, respectively. CONCLUSIONS: We demonstrate direct volumetric optical coherence elastography from phase image data. The approach does not rely on particular stimulation or sampling sequences and allows the estimation of elastic tissue properties of up to 40 Hz.

Needle tip force estimation by deep learning from raw spectral OCT data.

PURPOSE: Needle placement is a challenging problem for applications such as biopsy or brachytherapy. Tip force sensing can provide valuable feedback for needle navigation inside the tissue. For this purpose, fiber-optical sensors can be directly integrated into the needle tip. Optical coherence tomography (OCT) can be used to image tissue. Here, we study how to calibrate OCT to sense forces, e.g., during robotic needle placement. METHODS: We investigate whether using raw spectral OCT data without a typical image reconstruction can improve a deep learning-based calibration between optical signal and forces. For this purpose, we consider three different needles with a new, more robust design which are calibrated using convolutional neural networks (CNNs). We compare training the CNNs with the raw OCT signal and the reconstructed depth profiles. RESULTS: We find that using raw data as an input for the largest CNN model outperforms the use of reconstructed data with a mean absolute error of 5.81 mN compared to 8.04 mN. CONCLUSIONS: We find that deep learning with raw spectral OCT data can improve learning for the task of force estimation. Our needle design and calibration approach constitute a very accurate fiber-optical sensor for measuring forces at the needle tip.

Detection and Compensation of Periodic Motion in Magnetic Particle Imaging.

The temporal resolution of the tomographic imaging method magnetic particle imaging (MPI) is remarkably high. The spatial resolution is degraded for measured voltage signal with low signal-to-noise ratio, because the regularization in the image reconstruction step needs to be increased for system-matrix approaches and for deconvolution steps in x -space approaches. To improve the signal-to-noise ratio, blockwise averaging of the signal over time can be advantageous. However, since block-wise averaging decreases the temporal resolution, it prevents resolving the motion. In this paper, a framework for averaging motion-corrupted MPI raw data is proposed. The motion is considered to be periodic as it is the case for respiration and/or the heartbeat. The same state of motion is thus reached repeatedly in a time series exceeding the repetition time of the motion and can be used for averaging. As the motion process and the acquisition process are, in general, not synchronized, averaging of the captured MPI raw data corresponding to the same state of motion requires to shift the starting point of the individual frames. For high-frequency motion, a higher frame rate is potentially required. To address this issue, a binning method for using only parts of complete frames from a motion cycle is proposed that further reduces the motion artifacts in the final images. The frequency of motion is derived directly from the MPI raw data signal without the need to capture an additional navigator signal. Using a motion phantom, it is shown that the proposed method is capable of averaging experimental data with reduced motion artifacts. The methods are further validated on in-vivo data from mouse experiments to compensate the heartbeat.

Multivariate respiratory motion prediction.

In extracranial robotic radiotherapy, tumour motion is compensated by tracking external and internal surrogates. To compensate system specific time delays, time series prediction of the external optical surrogates is used. We investigate whether the prediction accuracy can be increased by expanding the current clinical setup by an accelerometer, a strain belt and a flow sensor. Four previously published prediction algorithms are adapted to multivariate inputs-normalized least mean squares (nLMS), wavelet-based least mean squares (wLMS), support vector regression (SVR) and relevance vector machines (RVM)-and evaluated for three different prediction horizons. The measurement involves 18 subjects and consists of two phases, focusing on long term trends (M1) and breathing artefacts (M2). To select the most relevant and least redundant sensors, a sequential forward selection (SFS) method is proposed. Using a multivariate setting, the results show that the clinically used nLMS algorithm is susceptible to large outliers. In the case of irregular breathing (M2), the mean root mean square error (RMSE) of a univariate nLMS algorithm is 0.66 mm and can be decreased to 0.46 mm by a multivariate RVM model (best algorithm on average). To investigate the full potential of this approach, the optimal sensor combination was also estimated on the complete test set. The results indicate that a further decrease in RMSE is possible for RVM (to 0.42 mm). This motivates further research about sensor selection methods. Besides the optical surrogates, the sensors most frequently selected by the algorithms are the accelerometer and the strain belt. These sensors could be easily integrated in the current clinical setup and would allow a more precise motion compensation.

Investigating recurrent neural networks for OCT A-scan based tissue analysis.

OBJECTIVES: Optical Coherence Tomography (OCT) has been proposed as a high resolution image modality to guide transbronchial biopsies. In this study we address the question, whether individual A-scans obtained in needle direction can contribute to the identification of pulmonary nodules. METHODS: OCT A-scans from freshly resected human lung tissue specimen were recorded through a customized needle with an embedded optical fiber. Bidirectional Long Short Term Memory networks (BLSTMs) were trained on randomly distributed training and test sets of the acquired A-scans. Patient specific training and different pre-processing steps were evaluated. RESULTS: Classification rates from 67.5% up to 76% were archived for different training scenarios. Sensitivity and specificity were highest for a patient specific training with 0.87 and 0.85. Low pass filtering decreased the accuracy from 73.2% on a reference distribution to 62.2% for higher cutoff frequencies and to 56% for lower cutoff frequencies. CONCLUSION: The results indicate that a grey value based classification is feasible and may provide additional information for diagnosis and navigation. Furthermore, the experiments show patient specific signal properties and indicate that the lower and upper parts of the frequency spectrum contribute to the classification.

Evaluating and comparing algorithms for respiratory motion prediction.

In robotic radiosurgery, it is necessary to compensate for systematic latencies arising from target tracking and mechanical constraints. This compensation is usually achieved by means of an algorithm which computes the future target position. In most scientific works on respiratory motion prediction, only one or two algorithms are evaluated on a limited amount of very short motion traces. The purpose of this work is to gain more insight into the real world capabilities of respiratory motion prediction methods by evaluating many algorithms on an unprecedented amount of data. We have evaluated six algorithms, the normalized least mean squares (nLMS), recursive least squares (RLS), multi-step linear methods (MULIN), wavelet-based multiscale autoregression (wLMS), extended Kalman filtering, and ε-support vector regression (SVRpred) methods, on an extensive database of 304 respiratory motion traces. The traces were collected during treatment with the CyberKnife (Accuray, Inc., Sunnyvale, CA, USA) and feature an average length of 71 min. Evaluation was done using a graphical prediction toolkit, which is available to the general public, as is the data we used. The experiments show that the nLMS algorithm-which is one of the algorithms currently used in the CyberKnife-is outperformed by all other methods. This is especially true in the case of the wLMS, the SVRpred, and the MULIN algorithms, which perform much better. The nLMS algorithm produces a relative root mean square (RMS) error of 75% or less (i.e., a reduction in error of 25% or more when compared to not doing prediction) in only 38% of the test cases, whereas the MULIN and SVRpred methods reach this level in more than 77%, the wLMS algorithm in more than 84% of the test cases. Our work shows that the wLMS algorithm is the most accurate algorithm and does not require parameter tuning, making it an ideal candidate for clinical implementation. Additionally, we have seen that the structure of a patient's respiratory motion trace has strong influence on the outcome of prediction. Further work is needed to determine a priori the suitability of an individual's respiratory behaviour to motion prediction.

Automatic scanning of large tissue areas in neurosurgery using optical coherence tomography.

BACKGROUND: With its high spatial and temporal resolution, optical coherence tomography (OCT) is an ideal modality for intra-operative imaging. One possible application is to detect tumour invaded tissue in neurosurgery, e.g. during complete resection of glioblastoma. Ideally, the whole resection cavity is scanned. However, OCT is limited to a small field of view (FOV) and scanning perpendicular to the tissue surface. METHODS: We present a new method to use OCT for scanning of the resection cavity during neurosurgical resection of brain tumours. The main challenges are creating a map of the cavity, scanning perpendicular to the surface and merging the three-dimensional (3D) data for intra-operative visualization and detection of residual tumour cells. RESULTS: Our results indicate that the proposed method enables creating high-resolution maps of the resection cavity. An overlay of these maps with the microscope images provides the surgeon with important information on the location of residual tumour tissue underneath the surface. CONCLUSION: We demonstrated that it is possible to automatically acquire an OCT image of the complete resection cavity. Overlaying microscopy images with depth information from OCT could lead to improved detection of residual tumour cells.

SU-E-T-618: Error Compensated Sparse Optimization for Fast Radiosurgery Treatment Planning.

PURPOSE: Radiosurgical treatment planning requires a good approximation of the dose distribution which is typically computed on a high resolution grid. However, the resulting optimization problem is large, and leads to substantial runtime. We study a sparse grid approach, for which we estimate and compensate for the expected deviations from the bounds. METHODS: We buildup an estimate of the hotspot error distribution by measuring the maximum dose deviation within a voxel for a large number of randomly generated beam configurations. This results in a conservative estimation of overdosage as a function of upper bound reduction for different grid sizes. We adjust the bounds for voxels inside the target volume (PTV) according to our estimation thus maintaining the likelihood of dose deviations within acceptable limits. The approach was applied to a prostate case, where the volumes of interest are large and close to each other. Our planning objective is a prescribed dose of 36.25 Gy to the 87% isodose. We employed constrained optimization to optimize the lower PTV bound on 2, 4, and 8mm isotropic grids. Results were computed on 1mm grid. RESULTS: The initial coverage was 93.7%, 92%, and 91%, and the volume exceeding the upper bound was 0.74%, 1.71%, and 9% for grid sizes of 2, 4, and 8mm, respectively. Changing the upper bound by 0.5% and 2.5% for the 4 and 8 mm grids resulted in only 0.75% and 2.2% of the volume exceeding the bound. The coverage did not change. Mean optimization times were 141.1, 22.6 and 3.4 minutes using the 2, 4 or 8mm grid, respectively. CONCLUSIONS: Experiments show that planning on a sparse grid can achieve comparable results with those of a high resolution grid, as long as the bounds are carefully balanced. This leads to substantially lower optimization times which facilitates interactive planning. This work was supported by the Graduate School for Computing in Medicine and Life Sciences funded by Germany’s Excellence Initiative [DFG GSC 235/1].

The design, physical properties and clinical utility of an iris collimator for robotic radiosurgery.

Robotic radiosurgery using more than one circular collimator can improve treatment plan quality and reduce total monitor units (MU). The rationale for an iris collimator that allows the field size to be varied during treatment delivery is to enable the benefits of multiple-field-size treatments to be realized with no increase in treatment time due to collimator exchange or multiple traversals of the robotic manipulator by allowing each beam to be delivered with any desired field size during a single traversal. This paper describes the Iris variable aperture collimator (Accuray Incorporated, Sunnyvale, CA, USA), which incorporates 12 tungsten-copper alloy segments in two banks of six. The banks are rotated by 30 degrees with respect to each other, which limits the radiation leakage between the collimator segments and produces a 12-sided polygonal treatment beam. The beam is approximately circular, with a root-mean-square (rms) deviation in the 50% dose radius of <0.8% (corresponding to <0.25 mm at the 60 mm field size) and an rms variation in the 20-80% penumbra width of about 0.1 mm at the 5 mm field size increasing to about 0.5 mm at 60 mm. The maximum measured collimator leakage dose rate was 0.07%. A commissioning method is described by which the average dose profile can be obtained from four profile measurements at each depth based on the periodicity of the isodose line variations with azimuthal angle. The penumbra of averaged profiles increased with field size and was typically 0.2-0.6 mm larger than that of an equivalent fixed circular collimator. The aperture reproducibility is < or =0.1 mm at the lower bank, diverging to < or =0.2 mm at a nominal treatment distance of 800 mm from the beam focus. Output factors (OFs) and tissue-phantom-ratio data are identical to those used for fixed collimators, except the OFs for the two smallest field sizes (5 and 7.5 mm) are considerably lower for the Iris Collimator. If average collimator profiles are used, the assumption of circular symmetry results in dose calculation errors that are <1 mm or <1% for single beams across the full range of field sizes; errors for multiple non-coplanar beam treatment plans are expected to be smaller. Treatment plans were generated for 19 cases using the Iris Collimator (12 field sizes) and also using one and three fixed collimators. The results of the treatment planning study demonstrate that the use of multiple field sizes achieves multiple plan quality improvements, including reduction of total MU, increase of target volume coverage and improvements in conformality and homogeneity compared with using a single field size for a large proportion of the cases studied. The Iris Collimator offers the potential to greatly increase the clinical application of multiple field sizes for robotic radiosurgery.

Stepwise multi-criteria optimization for robotic radiosurgery.

Achieving good conformality and a steep dose gradient around the target volume remains a key aspect of radiosurgery. Clearly, this involves a trade-off between target coverage, conformality of the dose distribution, and sparing of critical structures. Yet, image guidance and robotic beam placement have extended highly conformal dose delivery to extracranial and moving targets. Therefore, the multi-criteria nature of the optimization problem becomes even more apparent, as multiple conflicting clinical goals need to be considered coordinate to obtain an optimal treatment plan. Typically, planning for robotic radiosurgery is based on constrained optimization, namely linear programming. An extension of that approach is presented, such that each of the clinical goals can be addressed separately and in any sequential order. For a set of common clinical goals the mapping to a mathematical objective and a corresponding constraint is defined. The trade-off among the clinical goals is explored by modifying the constraints and optimizing a simple objective, while retaining feasibility of the solution. Moreover, it becomes immediately obvious whether a desired goal can be achieved and where a trade-off is possible. No importance factors or predefined prioritizations of clinical goals are necessary. The presented framework forms the basis for interactive and automated planning procedures. It is demonstrated for a sample case that the linear programming formulation is suitable to search for a clinically optimal treatment, and that the optimization steps can be performed quickly to establish that a Pareto-efficient solution has been found. Furthermore, it is demonstrated how the stepwise approach is preferable compared to modifying importance factors.

Technical description, phantom accuracy, and clinical feasibility for single-session lung radiosurgery using robotic image-guided real-time respiratory tumor tracking.

To describe the technological background, the accuracy, and clinical feasibility for single session lung radiosurgery using a real-time robotic system with respiratory tracking. The latest version of image-guided real-time respiratory tracking software (Synchrony, Accuray Incorporated, Sunnyvale, CA) was applied and is described. Accuracy measurements were performed using a newly designed moving phantom model. We treated 15 patients with 19 lung tumors with robotic radiosurgery (CyberKnife, Accuray) using the same treatment parameters for all patients. Ten patients had primary tumors and five had metastatic tumors. All patients underwent computed tomography-guided percutaneous placement of one fiducial directly into the tumor, and were all treated with single session radiosurgery to a dose of 24 Gy. Follow up CT scanning was performed every two months. All patients could be treated with the automated robotic technique. The respiratory tracking error was less than 1 mm and the overall shape of the dose profile was not affected by target motion and/or phase shift between fiducial and optical marker motion. Two patients required a chest tube insertion after fiducial implantation because of pneumothorax. One patient experienced nausea after treatment. No other short-term adverse reactions were found. One patient showed imaging signs of pneumonitis without a clinical correlation. Single-session radiosurgery for lung tumor tracking using the described technology is a stable, safe, and feasible concept for respiratory tracking of tumors during robotic lung radiosurgery in selected patients. Longer follow-up is needed for definitive clinical results.

Prospective, randomized comparison of laryngeal tube and laryngeal mask airway in pediatric patients.

BACKGROUND: While reports of the use of laryngeal mask airway (LMA)-Classic in great patient numbers are available, data on the use of the laryngeal tube (LT) in this age group is limited. The two devices are compared in a prospective randomized trial to evaluate success rates and quality of airway seal. METHODS: Sixty children, aged 2-8 years, scheduled for elective surgical interventions were randomized to be ventilated with LMA or LT. Standardized anesthesia was induced with fentanyl and propofol. Number of insertion attempts, time until first tidal volume and intraoperative tidal volumes, and peak pressures were recorded. Airway leak pressure was measured with cuff pressure adjusted to 60 cmH(2)O. RESULTS: Demographic data were comparable, average age in the LMA/LT group was 5.2 +/- 1.9/5.3 +/- 1.9 years. Insertion was successful in 29 of 30 patients in the LMA group (second attempt 8) and in all patients in the LT group (second attempt 3). Time until first tidal volume for LMA/LT was 23.1 +/- 7.3/19.2 +/- 8.6 s (P < 0.05). Peak airway pressures for LMA and LT were 15.3 +/- 3.4 and 17.1 +/- 4.0 cmH(2)O (P < 0.05) with tidal volumes of 10.2 +/- 2.2 and 10.2 +/- 1.9 ml.kg(-1), airway leak pressure was 19.2 +/- 8.6 cmH(2)O for LMA and 26.3 +/- 7.3 cmH(2)O for LT (P < 0.001). CONCLUSION: Insertion success rate is high with both LMA and LT in the age group studied. The airway leak pressure, serving as an estimate to judge quality of airway seal, is higher with the LT.

Accurate prediction of kidney allograft outcome based on creatinine course in the first 6 months posttransplant.

Most attempts to predict early kidney allograft loss are based on the patient and donor characteristics at baseline. We investigated how the early posttransplant creatinine course compares to baseline information in the prediction of kidney graft failure within the first 4 years after transplantation. Two approaches to create a prediction rule for early graft failure were evaluated. First, the whole data set was analysed using a decision-tree building software. The software, rpart, builds classification or regression models; the resulting models can be represented as binary trees. In the second approach, a Hill-Climbing algorithm was applied to define cut-off values for the median creatinine level and creatinine slope in the period between day 60 and 180 after transplantation. Of the 497 patients available for analysis, 52 (10.5%) experienced an early graft loss (graft loss within the first 4 years after transplantation). From the rpart algorithm, a single decision criterion emerged: Median creatinine value on days 60 to 180 higher than 3.1 mg/dL predicts early graft failure (accuracy 95.2% but sensitivity = 42.3%). In contrast, the Hill-Climbing algorithm delivered a cut-off of 1.8 mg/dL for the median creatinine level and a cut-off of 0.3 mg/dL per month for the creatinine slope (sensitivity = 69.5% and specificity 79.0%). Prediction rules based on median and slope of creatinine levels in the first half year after transplantation allow early identification of patients who are at risk of loosing their graft early after transplantation. These patients may benefit from therapeutic measures tailored for this high-risk setting.

Feasibility of four-dimensional conformal planning for robotic radiosurgery.

Organ motion can have a severe impact on the dose delivered by radiation therapy, and different procedures have been developed to address its effects. Conventional techniques include breath hold methods and gating. A different approach is the compensation for target motion by moving the treatment beams synchronously. Practical results have been reported for robot based radiosurgery, where a linear accelerator mounted on a robotic arm delivers the dose. However, not all organs move in the same way, which results in a relative motion of the beams with respect to the body and the tissues in the proximity of the tumor. This relative motion can severely effect the dose delivered to critical structures. We propose a method to incorporate motion in the treatment planning for robotic radiosurgery to avoid potential overdosing of organs surrounding the target. The method takes into account the motion of all considered volumes, which is discretized for dose calculations. Similarly, the beam motion is taken into account and the aggregated dose coefficient over all discrete steps is used for planning. We simulated the treatment of a moving target with three different planning methods. First, we computed beam weights based on a 3D planning situation and simulated treatment with organ motion and the beams moving synchronously to the target. Second, beam weights were computed by the 4D planning method incorporating the organ and beam motion and treatment was simulated for beams moving synchronously to the target. Third, the beam weights were determined by the 4D planning method with the beams fixed during planning and simulation. For comparison we also give results for the 3D treatment plan if there was no organ motion and when the plan is delivered by fixed beams in the presence of organ motion. The results indicate that the new 4D method is preferable and can further improve the overall conformality of motion compensated robotic radiosurgery.