S Lee

SU-F-R-40: Robustness Test of Computed Tomography Textures of Lung Tissues to Varying Scanning Protocols Using a Realistic Phantom Environment

PURPOSE The reliability of computed tomography (CT) textures is an important element of radiomics analysis. This study investigates the dependency of lung CT textures on different breathing phases and changes in CT image acquisition protocols in a realistic phantom setting. METHODS We investigated 11 CT texture features for radiation-induced lung disease from 3 categories (first-order, grey level co-ocurrence matrix (GLCM), and Laws filter). A biomechanical swine lung phantom was scanned at two breathing phases (inhale/exhale) and two scanning protocols set for PET/CT and diagnostic CT scanning. Lung volumes acquired from the CT images were divided into 2-dimensional sub-regions with a grid spacing of 31 mm. The distribution of the evaluated texture features from these sub-regions were compared between the two scanning protocols and two breathing phases. The significance of each factor on feature values were tested at 95% significance level using analysis of covariance (ANCOVA) model with interaction terms included. Robustness of a feature to a scanning factor was defined as non-significant dependence on the factor. RESULTS Three GLCM textures (variance, sum entropy, difference entropy) were robust to breathing changes. Two GLCM (variance, sum entropy) and 3 Laws filter textures (S5L5, E5L5, W5L5) were robust to scanner changes. Moreover, the two GLCM textures (variance, sum entropy) were consistent across all 4 scanning conditions. First-order features, especially Hounsfield unit intensity features, presented the most drastic variation up to 39%. CONCLUSION Amongst the studied features, GLCM and Laws filter texture features were more robust than first-order features. However, the majority of the features were modified by either breathing phase or scanner changes, suggesting a need for calibration when retrospectively comparing scans obtained at different conditions. Further investigation is necessary to identify the sensitivity of individual image acquisition parameters.

WE-E-BRE-05: Ensemble of Graphical Models for Predicting Radiation Pneumontis Risk.

PURPOSE We propose a prior knowledge-based approach to construct an interaction graph of biological and dosimetric radiation pneumontis (RP) covariates for the purpose of developing a RP risk classifier. METHODS We recruited 59 NSCLC patients who received curative radiotherapy with minimum 6 month follow-up. 16 RP events was observed (CTCAE grade ≥2). Blood serum was collected from every patient before (pre-RT) and during RT (mid-RT). From each sample the concentration of the following five candidate biomarkers were taken as covariates: alpha-2-macroglobulin (α2M), angiotensin converting enzyme (ACE), transforming growth factor β (TGF-β), interleukin-6 (IL-6), and osteopontin (OPN). Dose-volumetric parameters were also included as covariates. The number of biological and dosimetric covariates was reduced by a variable selection scheme implemented by L1-regularized logistic regression (LASSO). Posterior probability distribution of interaction graphs between the selected variables was estimated from the data under the literature-based prior knowledge to weight more heavily the graphs that contain the expected associations. A graph ensemble was formed by averaging the most probable graphs weighted by their posterior, creating a Bayesian Network (BN)-based RP risk classifier. RESULTS The LASSO selected the following 7 RP covariates: (1) pre-RT concentration level of α2M, (2) α2M level mid- RT/pre-RT, (3) pre-RT IL6 level, (4) IL6 level mid-RT/pre-RT, (5) ACE mid-RT/pre-RT, (6) PTV volume, and (7) mean lung dose (MLD). The ensemble BN model achieved the maximum sensitivity/specificity of 81%/84% and outperformed univariate dosimetric predictors as shown by larger AUC values (0.78∼0.81) compared with MLD (0.61), V20 (0.65) and V30 (0.70). The ensembles obtained by incorporating the prior knowledge improved classification performance for the ensemble size 5∼50. CONCLUSION We demonstrated a probabilistic ensemble method to detect robust associations between RP covariates and its potential to improve RP prediction accuracy. Our Bayesian approach to incorporate prior knowledge can enhance efficiency in searching of such associations from data. The authors acknowledge partial support by: 1) CREATE Medical Physics Research Training Network grant of the Natural Sciences and Engineering Research Council (Grant number: 432290) and 2) The Terry Fox Foundation Strategic Training Initiative for Excellence in Radiation Research for the 21st Century (EIRR21).

TU‐G‐108‐05: Assessment of Different Machine Learning Techniques for Multivariate Radiation Pneumonitis Modeling

PURPOSE We intended to perform an independent verification of different machine learning methods for inferring radiation pneumonitis (RP) risk from patient-specific biological and dosimetric informationMethods: 29 NSCLC patients who received chemoradiation were recruited from two institutions (22 from WUSTL, 7 from McGill). Blood samples were collected from each patient before and during radiotherapy (RT). From each sample the concentration of the five following candidate biomarkers were measured: alpha-2-macroglobulin (α2M), angiotensin converting enzyme (ACE), transforming growth factor β (TGF-β), interleukin-6 (IL-6), and osteopontin (OPN). Dimensionality of the raw biomarker data was reduced by a semi-supervised variable selection scheme. The reduced biomarker variables along with three known dosimetric RP predictors (mean lung dose, V20, tumor position in superior-inferior direction of the lung) were used as features for classifying high RP risk (CTCAE grade 2 or higher) patients. Four different machine learning methods (Bayesian Network, Logistic Regression, Naive Bayes, Support Vector Machine) were used for training classifiers on the WUSTL subset and were tested with respect to classification performance on the McGill subset. RESULTS The following 4 biomarker variables were chosen by the variable filtering: (1) pre-RT concentration level of α2M, (2) ratio of pre-to intra-RT levels of α2M, (3) intra-RT ACE level, (4) intra-RT TGFβ level. The best performance was achieved by Support Vector Machine in terms of classification accuracy (85.71%) and the area under the ROC curve (AUC) (0.9167). Bayesian Network recorded the same accuracy and slightly lower AUC (0.8750). Less accurate models were Naive Bayes and logistic regression in which overfitting was predominant with the largest difference in classification performance between the training and the testing dataset. CONCLUSION Preliminary results from this ongoing study suggests the need of a machine learning approach capable of modeling inter-relation between physical and biological variables where commonly used multivariate logistic regression would be inappropriate.

SU‐E‐T‐276: Treatment Planning Strategies for Lung Injury Studies in Rat Models in 6 MV Delivery

PURPOSE To study planning strategies that can be used in small animal radiation-induced lung toxicity experiments using 6 MV accelerator with high density MLC. METHODS Three different types of plans were designed on CT images of a Sprague Dawley rat model to irradiate 50% of the total lung volume (lung divided into apex and base) with a prescription dose of 24 Gy to the partial lung. Two VMAT arc therapy plans were optimized to cover to the prescription dose, either the apex or base of the lung. Two AP- PA plans were designed to completely block either lung apex or base while irradiating the remaining 50% of the lung. Finally, two AP-PA plans were designed to cover, to the prescription dose, the apex or base of the lung. The plans were designed and optimized using the Eclipse AAA algorithm and recalculated using the MMCTP/EGS/Beam Monte Carlo system. RESULTS When completely blocking the lung base, the apex will be underdosed by up to 30%; when completely covering the apex by the prescribed dose, the base will receive overdosing (V50%=73%). The VMAT plan leads to a more conformal dose distribution and spares unnecessary skin exposure when compared to AP-PA MV or kV delivery. Despite the small size of rat model, the 6 MV VMAT delivery is superior in terms of dose conformality and sparing of the heart and the non-irradiated 50% of the lung compared to the standard, simpler, AP-PA delivery. MC dosimetry in lung shows that the delivered dose is 10% higher than predicted by AAA because of the predominance of small fields in the delivery. CONCLUSIONS Clinical state-of- the-art planning and delivery techniques can be scaled down accurately to rats. The use of these techniques is essential in small animal studies to render conclusions of radiation response investigations translatable to human studies.

WE‐G‐BRA‐02: Model for Time‐Dependent Radiation‐Induced Lung Disease Risk Based on Systematic Image‐Based Scoring and Monte‐Carlo Dose Calculations

Purpose: To construct an analytical model for radiation‐induced lung injury (RILD) risk using computed tomography(CT)image analysis correlated to Monte Carlo (MC) dose calculations along with investigation of the models post‐RT dependency. Methods: The extent of RILD was segmented on the difference CTimage between a planning CTimage and registered post‐RT diagnostic CTimage. Radiation dose was calculated using the anisotropic analytical algorithm (AAA) and MC methods. The segmented RILD was spatially correlated with the dose distribution to generate a dose‐response curve for each of 39 follow‐up studies from 12 subjects. The response curves were grouped into 6 follow‐up periods with 3 months intervals according to the time elapsed since the completion of RT. For each period, a probit function derived from the Lyman‐Kutcher‐Burman (LKB) model was fit to the patient data with the two adjustable parameters: TD50 (dose at 50% chances of complication) and m (steepness of the curve). Results: TD50 demonstrated a monotonic increase from its initial level (73 Gy/77 Gy for AAA/MC dose) to its peak (130 Gy/116 Gy) at 9∼12 months post‐RT after which it fell to 85 Gy/80 Gy beyond 15 months post‐ RT. The change in TD50 occurred coincidently with the decrease in the proportion of injured lung volume, demonstrating the association between TD50 and the severity of RILD. The value of m significantly decreased in time from its initial values (0.51/0.55) to 0.24/0.25 beyond 15 months post‐RT. This suggests a transition in the dose‐response from a linear‐no‐threshold to nonlinear‐threshold type behavior. Replacement of AAA calculation by MC did not yield a significant difference in the fitting parameters. Conclusions: Time‐dependent results from the analytical modeling of RILD dose response indicates the transition from early to late radiation effects and the necessity to incorporate a temporal factor into the current time‐static RILD risk models.

Poster — Thur Eve — 46: Image-Based Scoring of Radiation Injury in Lung: Analysis of Sources of Uncertainties

We are studying the robustness and uncertainties of an automated method for quantifying radiotherapy‐induced lung injury from CTimages and delineate its relationship with radiationdose at different post‐radiation (post‐RT) time points. Methods: Using multi‐resolution affine optimization technique, post‐RT diagnostic CTimages were registered to planning CTimages. Following the registration and patient tissue‐based CTcalibration, a change in physical density at each voxel position of the planning CT was evaluated and voxels which density change is considered pathological were segmented as injury. Retrospective dose calculations using anisotropic analytical algorithm (AAA) and Monte‐Carlo (MC) were performed. The segmented injury was spatially correlated to the dose distributions to deduce a patient‐specific dose‐response relationship for radiation‐induced injury. Results: We found the probability of injury as a function of dose and post‐treatment time was patient‐specific. Due to the inaccuracy of the affine registration, the injury segmentation was manually corrected for the misalignment of normal tissue features, which gave rise to a case‐dependent uncertainty of up to 10%. Inter‐patient variability in CTcalibration contributed 4% or less to the uncertainty on the probability. Finally, dose calculation from MC simulation occasionally yielded a significantly modified complication probability compared to AAA model suggesting that dose calculation accuracy is important for the investigation on dose‐response of lung injury. Conclusion: The presented method provided a quantitative approach for dose‐responseanalysis in normal lungtissues if the accuracy in image registration and dose calculation can be assured and will provide better options to complication‐driven treatment planning.

WE‐C‐204B‐04: Image‐Based Scoring of Radiation Injury in Lung for Dose‐Effect Correlations: Analysis of Sources of Uncertainties

Purpose: To investigate the robustness and uncertainties of an automated method of quantifying radiation‐induced lung injuries (pneumonitis and fibrosis) from CTimage density changes and to correlate the probabilities for the injuries with radiation dose at different post‐radiation time points.Methods and Materials: Using multi‐resolution affine optimization technique, post‐radiotherapy (RT) diagnostic CTimages were registered to planning CTimages. Following the registration and patient tissue‐based CT number calibration, a change in physical density at each voxel position of the post‐RT CT was evaluated and the voxels which density change is considered pathological was segmented as injury. The PTV was excluded from the analysis due to the lack of functional information to differentiate between recurrence and injury. Retrospective patient dose calculations using the anisotropic analytical method (AAA) and the Monte‐Carlo (MC) were performed. The segmented injury was spatially correlated to the dose distributions to deduce a patient‐specific dose‐response relationship for the radiation‐induced injury. Results: Probability of lung injury as a function of dose and post‐treatment time is patient‐dependent and can be up to 70% at the highest dose. Due to the inaccuracy of the affine registration, the injury segmentation was manually corrected for the misalignment of normal tissue features, which gave rise to a case‐dependent uncertainty of up to 10%. Inter‐patient variability in CTcalibration contributed by 4% or less to the uncertainty on the complication probability and was dependent on dose. Finally, dose calculation from direct MC simulation occasionally yielded a significantly modified complication probability than using the AAA model suggesting that dose calculation accuracy is important in the accuracy of dose‐response model of lung.Conclusion: The presented method facilitates dose‐responseanalysis in normal lungtissues if the accuracy in image registration and dose calculation can be assured and will provide options to complication‐driven treatment planning.