Search NASA⌕ Search

SEARCH · Search NASA

Results for “Principal component analysis”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 901 records · Page 50

TH-E-BRF-07: Raman Spectroscopy for Radiation Treatment Response Assessment in a Lung Metastases Mouse Model

Purpose: Raman spectroscopy is an optical spectroscopic method used to probe chemical information about a target tissue. Our goal was to investigate whether Raman spectroscopy is able to distinguish lung tumors from normal lung tissue and whether this technique can identify the molecular changes induced by radiation. Methods: 4T1 mouse breast cancer cells were implanted subcutaneously into the flanks of 6 Balb/C female mice. Four additional mice were used as “normal lung” controls. After 14 days, 3 mice bearing tumors received 6Gy to the left lung with 6MV photons and the other three were treated as “unirradiated tumor” controls. At a 24-hour time point, lungs were excised and the specimens were sectioned using a cryostat; alternating sections were either stained with hematoxylin and eosin (H and E) for evaluation by a pathologist or unstained for Raman measurements. 240 total Raman spectra were collected; 84 from normal lung controls; 63 from unirradiated tumors and 64 from tumors irradiated with 6Gy in a single fraction. Raman spectra were also collected from normal lung tissues of mice with unirradiated tumors. Principal component analysis (PCA) and discriminant function analysis (DFA) were performed to analyze the data. Results: Raman bands assignable to DNA/RNA showed prominent contributions in tumor tissues while Raman bands associated with hemoglobin showed strong contributions in normal lung tissue. PCA/DFA analysis identified normal lung tissue and tumor with 100% and 98.4% accuracy, respectively, relative to pathologic scoring. Additionally, normal lung tissues from unirradiated mice bearing tumors were classified as normal with 100% accuracy. In a model consisting of unirradiated and irradiated tumors identification accuracy was 79.4% and 93.8% respectively, relative to pathologic assessment. Conclusion: Initial results demonstrate the promise for Raman spectroscopy in the diagnosis normal vs. lung metastases as well as the assessment of radiation therapy response. Further investigations are needed to validate these results. This study was supported in part by a research grant from Varian Medical Systems, Palo Alto, CA.

60 APPLIED LIFE SCIENCES↗

SU-C-BRF-03: PCA Modeling of Anatomical Changes During Head and Neck Radiation Therapy

Purpose: To develop principal component analysis (PCA) models from daily cone beam CTs (CBCTs) of head and neck (H and N) patients that could be used prospectively in adaptive radiation therapy (ART). Methods: : For 7 H and N patients, Pinnacle Treatment Planning System (Philips Healthcare) was used to retrospectively deformably register daily CBCTs to the planning CT. The number N of CBCTs per treatment course ranged from 14 to 22. For each patient a PCA model was built from the deformation vector fields (DVFs), after first subtracting the mean DVF, producing N eigen-DVFs (EDVFs). It was hypothesized that EDVFs with large eigenvalues represent the major anatomical deformations during the course of treatment, and that it is feasible to relate each EDVF to a clinically meaningful systematic or random change in anatomy, such as weight loss, neck flexion, etc. Results: DVFs contained on the order of 3×87×87×58=1.3 million scalar values (3 times the number of voxels in the registered volume). The top 3 eigenvalues accounted for ∼90% of variance. Anatomical changes corresponding to an EDVF were evaluated by generating a synthetic DVF, and applying that DVF to the CT to produce a synthetic CBCT. For all patients, the EDVF for the largest eigenvalue was interpreted to model weight loss. The EDVF for other eigenvalues appeared to represented quasi-random fraction-to-fraction changes. Conclusion: The leading EDVFs from single-patient PCA models have tentatively been identified with weight loss changes during treatment. Other EDVFs are tentatively identified as quasi-random inter-fraction changes. Clean separation of systematic and random components may require further work. This work is expected to facilitate development of population-based PCA models that can be used to prospectively identify significant anatomical changes, such as weight loss, early in treatment, triggering replanning where beneficial.

60 APPLIED LIFE SCIENCES↗

An initial study on the estimation of time-varying volumetric treatment images and 3D tumor localization from single MV cine EPID images

Purpose: In this work the authors develop and investigate the feasibility of a method to estimate time-varying volumetric images from individual MV cine electronic portal image device (EPID) images. Methods: The authors adopt a two-step approach to time-varying volumetric image estimation from a single cine EPID image. In the first step, a patient-specific motion model is constructed from 4DCT. In the second step, parameters in the motion model are tuned according to the information in the EPID image. The patient-specific motion model is based on a compact representation of lung motion represented in displacement vector fields (DVFs). DVFs are calculated through deformable image registration (DIR) of a reference 4DCT phase image (typically peak-exhale) to a set of 4DCT images corresponding to different phases of a breathing cycle. The salient characteristics in the DVFs are captured in a compact representation through principal component analysis (PCA). PCA decouples the spatial and temporal components of the DVFs. Spatial information is represented in eigenvectors and the temporal information is represented by eigen-coefficients. To generate a new volumetric image, the eigen-coefficients are updated via cost function optimization based on digitally reconstructed radiographs and projection images. The updated eigen-coefficients are then multiplied with the eigenvectors to obtain updated DVFs that, in turn, give the volumetric image corresponding to the cine EPID image. Results: The algorithm was tested on (1) Eight digital eXtended CArdiac-Torso phantom datasets based on different irregular patient breathing patterns and (2) patient cine EPID images acquired during SBRT treatments. The root-mean-squared tumor localization error is (0.73 ± 0.63 mm) for the XCAT data and (0.90 ± 0.65 mm) for the patient data. Conclusions: The authors introduced a novel method of estimating volumetric time-varying images from single cine EPID images and a PCA-based lung motion model. This is the first method to estimate volumetric time-varying images from single MV cine EPID images, and has the potential to provide volumetric information with no additional imaging dose to the patient.

60 APPLIED LIFE SCIENCES↗

Sci—Fri AM: Mountain — 04: Label-free Raman spectroscopy of single tumour cells detects early radiation-induced glycogen synthesis associated with increased radiation resistance

Purpose: To use label-free Raman spectroscopy (RS) for early treatment monitoring of tumour cell radioresistance. Methods: Three human tumour cell lines, two radioresistant (H460, SF{sub 2} = 0.57 and MCF7, SF{sub 2} = 0.70) and one radiosensitive (LNCaP, SF{sub 2} = 0.36), were irradiated with single fractions of 2, 4, 6, 8 or 10 Gy. In additional experiments, H460 and MCF7 cells were irradiated under co-treatment with the anti-diabetic drug metformin, a known radiosensitizing agent. Treated and control cultures were analyzed with RS daily for 3 days post-treatment. Single-cell Raman spectra were acquired from 20 live cells per sample, and experiments were repeated in triplicate. The combined data sets were analyzed with principal component analysis using standard algorithms. Cells from each culture were also subjected to standard assays for viability, proliferation, cell cycle, and radiation clonogenic survival. Results: The radioresistant cells (H460, MCF7) exhibited a RS molecular radiation response signature, detectable as early as 1 day post-treatment, of which radiation-induced glycogen synthesis is a significant contributor. The radiosensitive cells (LNCaP) exhibited negligible glycogen synthesis. Co-treatment with metformin in MCF7 cells blocked glycogen synthesis, reduced viability and proliferation, and increased radiosensitivity. Conversely, metformin co-treatment in H460 cells did not produce these same effects; importantly, both radiation-induced synthesis of glycogen and radiosensitivity were unaffected. Conclusions: Label-free RS can detect early glycogen synthesis post-irradiation, a previously undocumented metabolic mechanism associated with tumour cell radioresistance that can be targeted to increase radiosensitivity. RS monitoring of intratumoral glycogen may provide new opportunities for personalized combined modality radiotherapy treatments.

60 APPLIED LIFE SCIENCES↗

A method for volumetric imaging in radiotherapy using single x-ray projection

Purpose: It is an intriguing problem to generate an instantaneous volumetric image based on the corresponding x-ray projection. The purpose of this study is to develop a new method to achieve this goal via a sparse learning approach. Methods: To extract motion information hidden in projection images, the authors partitioned a projection image into small rectangular patches. The authors utilized a sparse learning method to automatically select patches that have a high correlation with principal component analysis (PCA) coefficients of a lung motion model. A model that maps the patch intensity to the PCA coefficients was built along with the patch selection process. Based on this model, a measured projection can be used to predict the PCA coefficients, which are then further used to generate a motion vector field and hence a volumetric image. The authors have also proposed an intensity baseline correction method based on the partitioned projection, in which the first and the second moments of pixel intensities at a patch in a simulated projection image are matched with those in a measured one via a linear transformation. The proposed method has been validated in both simulated data and real phantom data. Results: The algorithm is able to identify patches that contain relevant motion information such as the diaphragm region. It is found that an intensity baseline correction step is important to remove the systematic error in the motion prediction. For the simulation case, the sparse learning model reduced the prediction error for the first PCA coefficient to 5%, compared to the 10% error when sparse learning was not used, and the 95th percentile error for the predicted motion vector was reduced from 2.40 to 0.92 mm. In the phantom case with a regular tumor motion, the predicted tumor trajectory was successfully reconstructed with a 0.82 mm error for tumor center localization compared to a 1.66 mm error without using the sparse learning method. When the tumor motion was driven by a real patient breathing signal with irregular periods and amplitudes, the average tumor center error was 0.6 mm. The algorithm robustness with respect to sparsity level, patch size, and presence or absence of diaphragm, as well as computation time, has also been studied. Conclusions: The authors have developed a new method that automatically identifies motion information from an x-ray projection, based on which a volumetric image is generated.

60 APPLIED LIFE SCIENCES↗

SU-E-J-151: Day-To-Day Variations in Fraction-Specific Motion Modeling Using Patient 4DCBCT Images

Purpose: The goal of this study is to quantify the interfraction reproducibility of patient-specific motion models derived from 4DCBCT acquired on the day of treatment of lung cancer stereotactic body radiotherapy (SBRT) patients. Methods: Motion models are derived from patient 4DCBCT images acquired daily over 3–5 fractions of treatment by 1) applying deformable image registration between each 4DCBCT image and a reference phase from that day, resulting in a set of displacement vector fields (DVFs), and 2) performing principal component analysis (PCA) on the DVFs to derive a motion model. The motion model from the first day of treatment is compared to motion models from each successive day of treatment to quantify variability in motion models generated from different days. Four SBRT patient datasets have been acquired thus far in this IRB approved study. Results: Fraction-specific motion models for each fraction and patient were derived and PCA eigenvectors and their associated eigenvalues are compared for each fraction. For the first patient dataset, the average root mean square error between the first two eigenvectors associated with the highest two eigenvalues, in four fractions was 0.1, while it was 0.25 between the last three PCA eigenvectors associated with the lowest three eigenvalues. It was found that the eigenvectors and eigenvalues of PCA motion models for each treatment fraction have variations and the first few eigenvectors are shown to be more stable across treatment fractions than others. Conclusion: Analysis of this dataset showed that the first two eigenvectors of the PCA patient-specific motion models derived from 4DCBCT were stable over the course of several treatment fractions. The third, fourth, and fifth eigenvectors had larger variations.

60 APPLIED LIFE SCIENCES↗

SU-E-J-257: A PCA Model to Predict Adaptive Changes for Head&neck Patients Based On Extraction of Geometric Features From Daily CBCT Datasets

Purpose: Using daily cone beam CTs (CBCTs) to develop principal component analysis (PCA) models of anatomical changes in head and neck (H&N) patients and to assess the possibility of using these prospectively in adaptive radiation therapy (ART). Methods: Planning CT (pCT) images of 4 H&N patients were deformed to model several different systematic changes in patient anatomy during the course of the radiation therapy (RT). A Pinnacle plugin was used to linearly interpolate the systematic change in patient for the 35 fraction RT course and to generate a set of 35 synthetic CBCTs. Each synthetic CBCT represents the systematic change in patient anatomy for each fraction. Deformation vector fields (DVFs) were acquired between the pCT and synthetic CBCTs with random fraction-to-fraction changes were superimposed on the DVFs. A patient-specific PCA model was built using these DVFs containing systematic plus random changes. It was hypothesized that resulting eigenDVFs (EDVFs) with largest eigenvalues represent the major anatomical deformations during the course of treatment. Results: For all 4 patients, the PCA model provided different results depending on the type and size of systematic change in patient’s body. PCA was more successful in capturing the systematic changes early in the treatment course when these were of a larger scale with respect to the random fraction-to-fraction changes in patient’s anatomy. For smaller scale systematic changes, random changes in patient could completely “hide” the systematic change. Conclusion: The leading EDVF from the patientspecific PCA models could tentatively be identified as a major systematic change during treatment if the systematic change is large enough with respect to random fraction-to-fraction changes. Otherwise, leading EDVF could not represent systematic changes reliably. This work is expected to facilitate development of population-based PCA models that can be used to prospectively identify significant anatomical changes early in treatment. This work is supported in part by a grant from Varian Medical Systems, Palo Alto, CA.

60 APPLIED LIFE SCIENCES↗

MO-F-CAMPUS-J-02: Automatic Recognition of Patient Treatment Site in Portal Images Using Machine Learning

Purpose: To investigate the method to automatically recognize the treatment site in the X-Ray portal images. It could be useful to detect potential treatment errors, and to provide guidance to sequential tasks, e.g. automatically verify the patient daily setup. Methods: The portal images were exported from MOSAIQ as DICOM files, and were 1) processed with a threshold based intensity transformation algorithm to enhance contrast, and 2) where then down-sampled (from 1024×768 to 128×96) by using bi-cubic interpolation algorithm. An appearance-based vector space model (VSM) was used to rearrange the images into vectors. A principal component analysis (PCA) method was used to reduce the vector dimensions. A multi-class support vector machine (SVM), with radial basis function kernel, was used to build the treatment site recognition models. These models were then used to recognize the treatment sites in the portal image. Portal images of 120 patients were included in the study. The images were selected to cover six treatment sites: brain, head and neck, breast, lung, abdomen and pelvis. Each site had images of the twenty patients. Cross-validation experiments were performed to evaluate the performance. Results: MATLAB image processing Toolbox and scikit-learn (a machine learning library in python) were used to implement the proposed method. The average accuracies using the AP and RT images separately were 95% and 94% respectively. The average accuracy using AP and RT images together was 98%. Computation time was ∼0.16 seconds per patient with AP or RT image, ∼0.33 seconds per patient with both of AP and RT images. Conclusion: The proposed method of treatment site recognition is efficient and accurate. It is not sensitive to the differences of image intensity, size and positions of patients in the portal images. It could be useful for the patient safety assurance. The work was partially supported by a research grant from Varian Medical System.

60 APPLIED LIFE SCIENCES↗

WE-G-BRD-04: BEST IN PHYSICS (JOINT IMAGING-THERAPY): An Integrated Model-Based Intrafractional Organ Motion Tracking Approach with Dynamic MRI in Head and Neck Radiotherapy

Purpose: In-treatment dynamic cine images, provided by the first commercially available MRI-guided radiotherapy system, allow physicians to observe intrafractional motion of head and neck (H&N) internal structures. Nevertheless, high anatomical complexity and relatively poor cine image contrast/resolution have complicated automatic intrafractional motion evaluation. We proposed an integrated model-based approach to automatically delineate and analyze moving structures from on-board cine images. Methods: The H&N upper airway, a complex and highly deformable region wherein severe internal motion often occurs, was selected as the target-to-be-tracked. To reliably capture its motion, a hierarchical structure model containing three statistical shapes (face, face-jaw, and face-jaw-palate) was first built from a set of manually delineated shapes using principal component analysis. An integrated model-fitting algorithm was then employed to align the statistical shapes to the first to-be-detected cine frame, and multi-feature level-set contour propagation was performed to identify the airway shape change in the remaining frames. Ninety sagittal cine MR image sets, acquired from three H&N cancer patients, were utilized to demonstrate this approach. Results: The tracking accuracy was validated by comparing the results to the average of two manual delineations in 20 randomly selected images from each patient. The resulting dice similarity coefficient (93.28+/−1.46 %) and margin error (0.49+/−0.12 mm) showed good agreement with the manual results. Intrafractional displacements of anterior, posterior, inferior, and superior airway boundaries were observed, with values of 2.62+/−2.92, 1.78+/−1.43, 3.51+/−3.99, and 0.68+/−0.89 mm, respectively. The H&N airway motion was found to vary across directions, fractions, and patients, and highly correlated with patients’ respiratory frequency. Conclusion: We proposed the integrated computational approach, which for the first time allows to automatically identify the H&N upper airway and quantify in-treatment H&N internal motion in real-time. This approach can be applied to track other structures’ motion, and provide guidance on patient-specific prediction of intra-/inter-fractional structure displacements.

60 APPLIED LIFE SCIENCES↗

WE-G-BRD-06: Volumetric Cine MRI (VC-MRI) Estimated Based On Prior Knowledge for On-Board Target Localization

Purpose: To develop a technique to generate on-board VC-MRI using patient prior 4D-MRI, motion modeling and on-board 2D-cine MRI for real-time 3D target verification of liver and lung radiotherapy. Methods: The end-expiration phase images of a 4D-MRI acquired during patient simulation are used as patient prior images. Principal component analysis (PCA) is used to extract 3 major respiratory deformation patterns from the Deformation Field Maps (DFMs) generated between end-expiration phase and all other phases. On-board 2D-cine MRI images are acquired in the axial view. The on-board VC-MRI at any instant is considered as a deformation of the prior MRI at the end-expiration phase. The DFM is represented as a linear combination of the 3 major deformation patterns. The coefficients of the deformation patterns are solved by matching the corresponding 2D slice of the estimated VC-MRI with the acquired single 2D-cine MRI. The method was evaluated using both XCAT (a computerized patient model) simulation of lung cancer patients and MRI data from a real liver cancer patient. The 3D-MRI at every phase except end-expiration phase was used to simulate the ground-truth on-board VC-MRI at different instances, and the center-tumor slice was selected to simulate the on-board 2D-cine images. Results: Image subtraction of ground truth with estimated on-board VC-MRI shows fewer differences than image subtraction of ground truth with prior image. Excellent agreement between profiles was achieved. The normalized cross correlation coefficients between the estimated and ground-truth in the axial, coronal and sagittal views for each time step were >= 0.982, 0.905, 0.961 for XCAT data and >= 0.998, 0.911, 0.9541 for patient data. For XCAT data, the maximum-Volume-Percent-Difference between ground-truth and estimated tumor volumes was 1.6% and the maximum-Center-of-Mass-Shift was 0.9 mm. Conclusion: Preliminary studies demonstrated the feasibility to estimate real-time VC-MRI for on-board target localization before or during radiotherapy treatments. National Institutes of Health Grant No. R01-CA184173; Varian Medical System.

60 APPLIED LIFE SCIENCES↗

WE-G-207-06: 3D Fluoroscopic Image Generation From Patient-Specific 4DCBCT-Based Motion Models Derived From Physical Phantom and Clinical Patient Images

Purpose: Respiratory-correlated cone-beam CT (4DCBCT) images acquired immediately prior to treatment have the potential to represent patient motion patterns and anatomy during treatment, including both intra- and inter-fractional changes. We develop a method to generate patient-specific motion models based on 4DCBCT images acquired with existing clinical equipment and used to generate time varying volumetric images (3D fluoroscopic images) representing motion during treatment delivery. Methods: Motion models are derived by deformably registering each 4DCBCT phase to a reference phase, and performing principal component analysis (PCA) on the resulting displacement vector fields. 3D fluoroscopic images are estimated by optimizing the resulting PCA coefficients iteratively through comparison of the cone-beam projections simulating kV treatment imaging and digitally reconstructed radiographs generated from the motion model. Patient and physical phantom datasets are used to evaluate the method in terms of tumor localization error compared to manually defined ground truth positions. Results: 4DCBCT-based motion models were derived and used to generate 3D fluoroscopic images at treatment time. For the patient datasets, the average tumor localization error and the 95th percentile were 1.57 and 3.13 respectively in subsets of four patient datasets. For the physical phantom datasets, the average tumor localization error and the 95th percentile were 1.14 and 2.78 respectively in two datasets. 4DCBCT motion models are shown to perform well in the context of generating 3D fluoroscopic images due to their ability to reproduce anatomical changes at treatment time. Conclusion: This study showed the feasibility of deriving 4DCBCT-based motion models and using them to generate 3D fluoroscopic images at treatment time in real clinical settings. 4DCBCT-based motion models were found to account for the 3D non-rigid motion of the patient anatomy during treatment and have the potential to localize tumor and other patient anatomical structures at treatment time even when inter-fractional changes occur. This project was supported, in part, through a Master Research Agreement with Varian Medical Systems, Inc., Palo Alto, CA. The project was also supported, in part, by Award Number R21CA156068 from the National Cancer Institute.

60 APPLIED LIFE SCIENCES↗

Benefits of adaptive radiation therapy in lung cancer as a function of replanning frequency

Purpose: To quantify the potential benefit associated with daily replanning in lung cancer in terms of normal tissue dose sparing and to characterize the tradeoff between adaptive benefit and replanning frequency. Methods: A set of synthetic images and contours, derived from weekly active breathing control images of 12 patients who underwent radiation therapy treatment for nonsmall cell lung cancer, is generated for each fraction of treatment using principal component analysis in a way that preserves temporal anatomical trends (e.g., tumor regression). Daily synthetic images and contours are used to simulate four different treatment scenarios: (1) a “no-adapt” scenario that simulates delivery of an initial plan throughout treatment, (2) a “midadapt” scenario that implements a single replan for fraction 18, (3) a “weekly adapt” scenario that simulates weekly adaptations, and (4) a “full-adapt” scenario that simulates daily replanning. An initial intensity modulated radiation therapy plan is created for each patient and replanning is carried out in an automated fashion by reoptimizing beam apertures and weights. Dose is calculated on each image and accumulated to the first in the series using deformable mappings utilized in synthetic image creation for comparison between simulated treatments. Results: Target coverage was maintained and cord tolerance was not exceeded for any of the adaptive simulations. Average reductions in mean lung dose (MLD) and volume of lung receiving 20 Gy or more (V20{sub lung}) were 65 ± 49 cGy (p = 0.000 01) and 1.1% ± 1.2% (p = 0.0006), respectively, for all patients. The largest reduction in MLD for a single patient was 162 cGy, which allowed an isotoxic escalation of the target dose of 1668 cGy. Average reductions in cord max dose, mean esophageal dose (MED), dose received by 66% of the heart (D66{sub heart}), and dose received by 33% of the heart (D33{sub heart}), were 158 ± 280, 117 ± 121, 37 ± 77, and 99 ± 120 cGy, respectively. Average incremental reductions in MLD for the midadapt, weekly adapt, and full-adapt treatments were 38, 18, and 8 cGy, respectively. Incremental reductions in MED for the same treatments were 57, 37, and 23 cGy. Reductions in MLD and MED for the full-adapt treatment were correlated with the absolute decrease in the planning target volume (r = 0.34 and r = 0.26). Conclusions: Adaptive radiation therapy for lung cancer yields clinically relevant reductions in normal tissue doses for frequencies of adaptation ranging from a single replan up to daily replanning. Increased frequencies of adaptation result in additional benefit while magnitude of benefit decreases.

60 APPLIED LIFE SCIENCES↗

SU-C-BRA-07: Variability of Patient-Specific Motion Models Derived Using Different Deformable Image Registration Algorithms for Lung Cancer Stereotactic Body Radiotherapy (SBRT) Patients

Purpose: To study the variability of patient-specific motion models derived from 4-dimensional CT (4DCT) images using different deformable image registration (DIR) algorithms for lung cancer stereotactic body radiotherapy (SBRT) patients. Methods: Motion models are derived by 1) applying DIR between each 4DCT image and a reference image, resulting in a set of displacement vector fields (DVFs), and 2) performing principal component analysis (PCA) on the DVFs, resulting in a motion model (a set of eigenvectors capturing the variations in the DVFs). Three DIR algorithms were used: 1) Demons, 2) Horn-Schunck, and 3) iterative optical flow. The motion models derived were compared using patient 4DCT scans. Results: Motion models were derived and the variations were evaluated according to three criteria: 1) the average root mean square (RMS) difference which measures the absolute difference between the components of the eigenvectors, 2) the dot product between the eigenvectors which measures the angular difference between the eigenvectors in space, and 3) the Euclidean Model Norm (EMN), which is calculated by summing the dot products of an eigenvector with the first three eigenvectors from the reference motion model in quadrature. EMN measures how well an eigenvector can be reconstructed using another motion model derived using a different DIR algorithm. Results showed that comparing to a reference motion model (derived using the Demons algorithm), the eigenvectors of the motion model derived using the iterative optical flow algorithm has smaller RMS, larger dot product, and larger EMN values than those of the motion model derived using Horn-Schunck algorithm. Conclusion: The study showed that motion models vary depending on which DIR algorithms were used to derive them. The choice of a DIR algorithm may affect the accuracy of the resulting model, and it is important to assess the suitability of the algorithm chosen for a particular application. This project was supported, in part, through a Master Research Agreement with Varian Medical Systems, Inc, Palo Alto, CA.

60 APPLIED LIFE SCIENCES↗

SU-C-209-02: 3D Fluoroscopic Image Generation From Patient-Specific 4DCBCT-Based Motion Models Derived From Clinical Patient Images

Purpose: We develop a method to generate time varying volumetric images (3D fluoroscopic images) using patient-specific motion models derived from four-dimensional cone-beam CT (4DCBCT). Methods: Motion models are derived by selecting one 4DCBCT phase as a reference image, and registering the remaining images to it. Principal component analysis (PCA) is performed on the resultant displacement vector fields (DVFs) to create a reduced set of PCA eigenvectors that capture the majority of respiratory motion. 3D fluoroscopic images are generated by optimizing the weights of the PCA eigenvectors iteratively through comparison of measured cone-beam projections and simulated projections generated from the motion model. This method was applied to images from five lung-cancer patients. The spatial accuracy of this method is evaluated by comparing landmark positions in the 3D fluoroscopic images to manually defined ground truth positions in the patient cone-beam projections. Results: 4DCBCT motion models were shown to accurately generate 3D fluoroscopic images when the patient cone-beam projections contained clearly visible structures moving with respiration (e.g., the diaphragm). When no moving anatomical structure was clearly visible in the projections, the 3D fluoroscopic images generated did not capture breathing deformations, and reverted to the reference image. For the subset of 3D fluoroscopic images generated from projections with visibly moving anatomy, the average tumor localization error and the 95th percentile were 1.6 mm and 3.1 mm respectively. Conclusion: This study showed that 4DCBCT-based 3D fluoroscopic images can accurately capture respiratory deformations in a patient dataset, so long as the cone-beam projections used contain visible structures that move with respiration. For clinical implementation of 3D fluoroscopic imaging for treatment verification, an imaging field of view (FOV) that contains visible structures moving with respiration should be selected. If no other appropriate structures are visible, the images should include the diaphragm. This project was supported, in part, through a Master Research Agreement with Varian Medical Systems, Inc, Palo Alto, CA.

60 APPLIED LIFE SCIENCES↗

SU-D-201-05: On the Automatic Recognition of Patient Safety Hazards in a Radiotherapy Setup Using a Novel 3D Camera System and a Deep Learning Framework

Purpose: Patient safety hazards such as a wrong patient/site getting treated can lead to catastrophic results. The purpose of this project is to automatically detect potential patient safety hazards during the radiotherapy setup and alert the therapist before the treatment is initiated. Methods: We employed a set of co-located and co-registered 3D cameras placed inside the treatment room. Each camera provided a point-cloud of fraxels (fragment pixels with 3D depth information). Each of the cameras were calibrated using a custom-built calibration target to provide 3D information with less than 2 mm error in the 500 mm neighborhood around the isocenter. To identify potential patient safety hazards, the treatment room components and the patient’s body needed to be identified and tracked in real-time. For feature recognition purposes, we used a graph-cut based feature recognition with principal component analysis (PCA) based feature-to-object correlation to segment the objects in real-time. Changes in the object’s position were tracked using the CamShift algorithm. The 3D object information was then stored for each classified object (e.g. gantry, couch). A deep learning framework was then used to analyze all the classified objects in both 2D and 3D and was then used to fine-tune a convolutional network for object recognition. The number of network layers were optimized to identify the tracked objects with >95% accuracy. Results: Our systematic analyses showed that, the system was effectively able to recognize wrong patient setups and wrong patient accessories. The combined usage of 2D camera information (color + depth) enabled a topology-preserving approach to verify patient safety hazards in an automatic manner and even in scenarios where the depth information is partially available. Conclusion: By utilizing the 3D cameras inside the treatment room and a deep learning based image classification, potential patient safety hazards can be effectively avoided.

60 APPLIED LIFE SCIENCES↗

SU-F-R-41: Regularized PCA Can Model Treatment-Related Changes in Head and Neck Patients Using Daily CBCTs

Purpose: To use daily cone beam CTs (CBCTs) to develop regularized principal component analysis (PCA) models of anatomical changes in head and neck (H&N) patients, to guide replanning decisions in adaptive radiation therapy (ART). Methods: Known deformations were applied to planning CT (pCT) images of 10 H&N patients to model several different systematic anatomical changes. A Pinnacle plugin was used to interpolate systematic changes over 35 fractions, generating a set of 35 synthetic CTs for each patient. Deformation vector fields (DVFs) were acquired between the pCT and synthetic CTs and random fraction-to-fraction changes were superimposed on the DVFs. Standard non-regularized and regularized patient-specific PCA models were built using the DVFs. The ability of PCA to extract the known deformations was quantified. PCA models were also generated from clinical CBCTs, for which the deformations and DVFs were not known. It was hypothesized that resulting eigenvectors/eigenfunctions with largest eigenvalues represent the major anatomical deformations during the course of treatment. Results: As demonstrated with quantitative results in the supporting document regularized PCA is more successful than standard PCA at capturing systematic changes early in the treatment. Regularized PCA is able to detect smaller systematic changes against the background of random fraction-to-fraction changes. To be successful at guiding ART, regularized PCA should be coupled with models of when anatomical changes occur: early, late or throughout the treatment course. Conclusion: The leading eigenvector/eigenfunction from the both PCA approaches can tentatively be identified as a major systematic change during radiotherapy course when systematic changes are large enough with respect to random fraction-to-fraction changes. In all cases the regularized PCA approach appears to be more reliable at capturing systematic changes, enabling dosimetric consequences to be projected once trends are established early in the treatment course. This work is supported in part by a grant from Varian Medical Systems, Palo Alto, CA.

60 APPLIED LIFE SCIENCES↗

SU-F-J-138: An Extension of PCA-Based Respiratory Deformation Modeling Via Multi-Linear Decomposition

Purpose: To address and lift the limited degree of freedom (DoF) of globally bilinear motion components such as those based on principal components analysis (PCA), for encoding and modeling volumetric deformation motion. Methods: We provide a systematic approach to obtaining a multi-linear decomposition (MLD) and associated motion model from deformation vector field (DVF) data. We had previously introduced MLD for capturing multi-way relationships between DVF variables, without being restricted by the bilinear component format of PCA-based models. PCA-based modeling is commonly used for encoding patient-specific deformation as per planning 4D-CT images, and aiding on-board motion estimation during radiotherapy. However, the bilinear space-time decomposition inherently limits the DoF of such models by the small number of respiratory phases. While this limit is not reached in model studies using analytical or digital phantoms with low-rank motion, it compromises modeling power in the presence of relative motion, asymmetries and hysteresis, etc, which are often observed in patient data. Specifically, a low-DoF model will spuriously couple incoherent motion components, compromising its adaptability to on-board deformation changes. By the multi-linear format of extracted motion components, MLD-based models can encode higher-DoF deformation structure. Results: We conduct mathematical and experimental comparisons between PCA- and MLD-based models. A set of temporally-sampled analytical trajectories provides a synthetic, high-rank DVF; trajectories correspond to respiratory and cardiac motion factors, including different relative frequencies and spatial variations. Additionally, a digital XCAT phantom is used to simulate a lung lesion deforming incoherently with respect to the body, which adheres to a simple respiratory trend. In both cases, coupling of incoherent motion components due to a low model DoF is clearly demonstrated. Conclusion: Multi-linear decomposition can enable decoupling of distinct motion factors in high-rank DVF measurements. This may improve motion model expressiveness and adaptability to on-board deformation, aiding model-based image reconstruction for target verification. NIH Grant No. R01-184173.

60 APPLIED LIFE SCIENCES↗

SU-G-BRC-13: Model Based Classification for Optimal Position Selection for Left-Sided Breast Radiotherapy: Free Breathing, DIBH, Or Prone

Purpose: There are clinical decision challenges to select optimal treatment positions for left-sided breast cancer patients—supine free breathing (FB), supine Deep Inspiration Breath Hold (DIBH) and prone free breathing (prone). Physicians often make the decision based on experiences and trials, which might not always result optimal OAR doses. We herein propose a mathematical model to predict the lowest OAR doses among these three positions, providing a quantitative tool for corresponding clinical decision. Methods: Patients were scanned in FB, DIBH, and prone positions under an IRB approved protocol. Tangential beam plans were generated for each position, and OAR doses were calculated. The position with least OAR doses is defined as the optimal position. The following features were extracted from each scan to build the model: heart, ipsilateral lung, breast volume, in-field heart, ipsilateral lung volume, distance between heart and target, laterality of heart, and dose to heart and ipsilateral lung. Principal Components Analysis (PCA) was applied to remove the co-linearity of the input data and also to lower the data dimensionality. Feature selection, another method to reduce dimensionality, was applied as a comparison. Support Vector Machine (SVM) was then used for classification. Thirtyseven patient data were acquired; up to now, five patient plans were available. K-fold cross validation was used to validate the accuracy of the classifier model with small training size. Results: The classification results and K-fold cross validation demonstrated the model is capable of predicting the optimal position for patients. The accuracy of K-fold cross validations has reached 80%. Compared to PCA, feature selection allows causal features of dose to be determined. This provides more clinical insights. Conclusion: The proposed classification system appeared to be feasible. We are generating plans for the rest of the 37 patient images, and more statistically significant results are to be presented.

60 APPLIED LIFE SCIENCES↗