Search NASA⌕ Search

SEARCH · Search NASA

Results for “FEATURE”

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 163 records · Page 9

[Fe II] 1.64 μm IMAGING OBSERVATIONS OF THE OUTFLOW FEATURES AROUND ULTRACOMPACT H II REGIONS IN THE FIRST GALACTIC QUADRANT

We present [Fe II] 1.644 μm features around ultracompact H II regions (UCHIIs) found on a quest for the ''footprint'' outflow features of UCHIIs—the features produced by outflowing materials ejected during an earlier, active accretion phase of massive young stellar objects (MYSOs). We surveyed 237 UCHIIs in the first Galactic quadrant, employing the CORNISH UCHII catalog and UWIFE data, which is an imaging survey in [Fe II] 1.644 μm performed with UKIRT-WFCAM under ∼0.''8 seeing conditions. The [Fe II] features were found around five UCHIIs, one of which was less plausible. We interpret the [Fe II] features to be shock-excited by outflows from YSOs and estimate the outflow mass-loss rates from the [Fe II] flux which are ∼1 × 10{sup –6}-4 × 10{sup –5} M {sub ☉} yr{sup –1}. We propose that the [Fe II] features might be the ''footprint'' outflow features, but more studies are required to clarify whether or not this is the case. This is based on the morphological relation between the [Fe II] and 5 GHz radio features, the outflow mass-loss rate, the travel time of the [Fe II] features, and the existence of several YSO candidates near the UCHIIs. The UCHIIs accompanying the [Fe II] features have relatively higher peak flux densities. The fraction of UCHIIs accompanying the [Fe II] features, 5/237, is small when compared to the ∼90% detection rate of high-velocity CO gas around UCHIIs. We discuss some possible explanations for the low detection rate.

79 ASTRONOMY AND ASTROPHYSICS↗

WinnowML: Stable feature selection for maximizing prediction accuracy of time-based system modeling

Online deep learning (ODL) has become an important methodology for modeling time-based performance of computer systems. An open problem is the intelligent selection of features from raw workload traces of computer systems. The best methods are overly sensitive to noisy data, causing frequent feature changes and re-training. Using all available features inflates training time and introduces model artifacts if some features should have been dropped. We present WinnowML, a method for automatically determining the most relevant feature subset for a predictive time-series model. WinnowML combines existing feature ranking algorithms and a history of each feature's ranking to iteratively rank a feature set to lower prediction error and maximize long term relevance. From this ranked feature set, the most relevant and stable subset is selected to train a model. Experimentally, we show how WinnowML can lower a model's mean absolute relative error up to 42% on average compared to the closest performing approach. Additionally, we lower the fluctuation in feature ranking and selection up to 65%. We also demonstrate how to combine WinnowML and a model search tool to provide improvements in performance of up to 14.5% when compared to using all the feature available.

Bel, Oceane MS↗

TU-C-12A-09: Modeling Pathologic Response of Locally Advanced Esophageal Cancer to Chemo-Radiotherapy Using Quantitative PET/CT Features, Clinical Parameters and Demographics

Purpose: To develop predictive models using quantitative PET/CT features for the evaluation of tumor response to neoadjuvant chemo-radiotherapy (CRT) in patients with locally advanced esophageal cancer. Methods: This study included 20 patients who underwent tri-modality therapy (CRT + surgery) and had {sup 18}F-FDG PET/CT scans before initiation of CRT and 4-6 weeks after completion of CRT but prior to surgery. Four groups of tumor features were examined: (1) conventional PET/CT response measures (SUVmax, tumor diameter, etc.); (2) clinical parameters (TNM stage, histology, etc.) and demographics; (3) spatial-temporal PET features, which characterize tumor SUV intensity distribution, spatial patterns, geometry, and associated changes resulting from CRT; and (4) all features combined. An optimal feature set was identified with recursive feature selection and cross-validations. Support vector machine (SVM) and logistic regression (LR) models were constructed for prediction of pathologic tumor response to CRT, using cross-validations to avoid model over-fitting. Prediction accuracy was assessed via area under the receiver operating characteristic curve (AUC), and precision was evaluated via confidence intervals (CIs) of AUC. Results: When applied to the 4 groups of tumor features, the LR model achieved AUCs (95% CI) of 0.57 (0.10), 0.73 (0.07), 0.90 (0.06), and 0.90 (0.06). The SVM model achieved AUCs (95% CI) of 0.56 (0.07), 0.60 (0.06), 0.94 (0.02), and 1.00 (no misclassifications). Using spatial-temporal PET features combined with conventional PET/CT measures and clinical parameters, the SVM model achieved very high accuracy (AUC 1.00) and precision (no misclassifications), significantly better than using conventional PET/CT measures or clinical parameters and demographics alone. For groups with a large number of tumor features (groups 3 and 4), the SVM model achieved significantly higher accuracy than the LR model. Conclusion: The SVM model using all features including quantitative PET/CT features accurately and precisely predicted pathologic tumor response to CRT in esophageal cancer. This work was supported in part by National Cancer Institute Grant R21 CA131979 and R01 CA172638. Shan Tan was supported in part by the National Natural Science Foundation of China 60971112 and 61375018, and by Fundamental Research Funds for the Central Universities 2012QN086.

60 APPLIED LIFE SCIENCES↗

SU-D-BRA-06: Dual-Energy Chest CT: The Effects of Virtual Monochromatic Reconstructions On Texture Analysis Features

Purpose: To characterize the effect of virtual monochromatic reconstructions on several commonly used texture analysis features in DECT of the chest. Further, to assess the effect of monochromatic energy levels on the ability of these textural features to identify tissue types. Methods: 20 consecutive patients underwent chest CTs for evaluation of lung nodules using Siemens Somatom Definition Flash DECT. Virtual monochromatic images were constructed at 10keV intervals from 40–190keV. For each patient, an ROI delineated the lesion under investigation, and cylindrical ROI’s were placed within 5 different healthy tissues (blood, fat, muscle, lung, and liver). Several histogram- and Grey Level Cooccurrence Matrix (GLCM)-based texture features were then evaluated in each ROI at each energy level. As a means of validation, these feature values were then used in a random forest classifier to attempt to identify the tissue types present within each ROI. Their predictive accuracy at each energy level was recorded. Results: All textural features changed considerably with virtual monochromatic energy, particularly below 70keV. Most features exhibited a global minimum or maximum around 80keV, and while feature values changed with energy above this, patient ranking was generally unaffected. As expected, blood demonstrated the lowest inter-patient variability, for all features, while lung lesions (encompassing many different pathologies) exhibited the highest. The accuracy of these features in identifying tissues (76% accuracy) was highest at 80keV, but no clear relationship between energy and classification accuracy was found. Two common misclassifications (blood vs liver and muscle vs fat) accounted for the majority (24 of the 28) errors observed. Conclusion: All textural features were highly dependent on virtual monochromatic energy level, especially below 80keV, and were more stable above this energy. However, in a random forest model, these commonly used features were able to reliably differentiate between most tissues types regardless of energy level. Dr Godoy has received a dual-energy CT research grant from Siemens Healthcare. That grant did not directly fund this research.

60 APPLIED LIFE SCIENCES↗

SU-D-BRA-07: A Phantom Study to Assess the Variability in Radiomics Features Extracted From Cone-Beam CT Images

Purpose: Several studies have demonstrated the prognostic potential for texture features extracted from CT images of non-small cell lung cancer (NSCLC) patients. The purpose of this study was to determine if these features could be extracted with high reproducibility from cone-beam CT (CBCT) images in order for features to be easily tracked throughout a patient’s treatment. Methods: Two materials in a radiomics phantom, designed to approximate NSCLC tumor texture, were used to assess the reproducibility of 26 features. This phantom was imaged on 9 CBCT scanners, including Elekta and Varian machines. Thoracic and head imaging protocols were acquired on each machine. CBCT images from 27 NSCLC patients imaged using the thoracic protocol on Varian machines were obtained for comparison. The variance for each texture measured from these patients was compared to the variance in phantom values for different manufacturer/protocol subsets. Levene’s test was used to identify features which had a significantly smaller variance in the phantom scans versus the patient data. Results: Approximately half of the features (13/26 for material1 and 15/26 for material2) had a significantly smaller variance (p<0.05) between Varian thoracic scans of the phantom compared to patient scans. Many of these same features remained significant for the head scans on Varian (12/26 and 8/26). However, when thoracic scans from Elekta and Varian were combined, only a few features were still significant (4/26 and 5/26). Three features (skewness, coarsely filtered mean and standard deviation) were significant in almost all manufacturer/protocol subsets. Conclusion: Texture features extracted from CBCT images of a radiomics phantom are reproducible and show significantly less variation than the same features measured from patient images when images from the same manufacturer or with similar parameters are used. Reproducibility between CBCT scanners may be high enough to allow the extraction of meaningful texture values for patients. This project was funded in part by the Cancer Prevention Research Institute of Texas (CPRIT). Xenia Fave is a recipient of the American Association of Physicists in Medicine Graduate Fellowship.

60 APPLIED LIFE SCIENCES↗

TU-AB-BRA-06: Texture Feature Reproducibility Between PET/CT and PET/MR Imaging Modalities

Purpose: Radiomics involves the extraction of texture features from different imaging modalities with the purpose of developing models to predict patient treatment outcomes. The purpose of this study is to investigate texture feature reproducibility across [18F]FDG PET/CT and [18F]FDG PET/MR imaging in patients with primary malignancies. Methods: Twenty five prospective patients with solid tumors underwent clinical [18F]FDG PET/CT scan followed by [18F]FDG PET/MR scans. In all patients the lesions were identified using nuclear medicine reports. The images were co-registered and segmented using an in-house auto-segmentation method. Fifty features, based on the intensity histogram, second and high order matrices, were extracted from the segmented regions from both image data sets. One-way random-effects ANOVA model of the intra-class correlation coefficient (ICC) was used to establish texture feature correlations between both data sets. Results: Fifty features were classified based on their ICC values, which were found in the range from 0.1 to 0.86, in three categories: high, intermediate, and low. Ten features extracted from second and high-order matrices showed large ICC ≥ 0.70. Seventeen features presented intermediate 0.5 ≤ ICC ≤ 0.65 and the remaining twenty three presented low ICC ≤ 0.45. Conclusion: Features with large ICC values could be reliable candidates for quantification as they lead to similar results from both imaging modalities. Features with small ICC indicates a lack of correlation. Therefore, the use of these features as a quantitative measure will lead to different assessments of the same lesion depending on the imaging modality from where they are extracted. This study shows the importance of the need for further investigation and standardization of features across multiple imaging modalities.

60 APPLIED LIFE SCIENCES↗

SU-F-R-08: Can Normalization of Brain MRI Texture Features Reduce Scanner-Dependent Effects in Unsupervised Machine Learning?

Purpose: To reduce differences in features calculated from MRI brain scans acquired at different field strengths with or without Gadolinium contrast. Methods: Brain scans were processed for 111 epilepsy patients to extract hippocampus and thalamus features. Scans were acquired on 1.5 T scanners with Gadolinium contrast (group A), 1.5T scanners without Gd (group B), and 3.0 T scanners without Gd (group C). A total of 72 features were extracted. Features were extracted from original scans and from scans where the image pixel values were rescaled to the mean of the hippocampi and thalami values. For each data set, cluster analysis was performed on the raw feature set and for feature sets with normalization (conversion to Z scores). Two methods of normalization were used: The first was over all values of a given feature, and the second by normalizing within the patient group membership. The clustering software was configured to produce 3 clusters. Group fractions in each cluster were calculated. Results: For features calculated from both the non-rescaled and rescaled data, cluster membership was identical for both the non-normalized and normalized data sets. Cluster 1 was comprised entirely of Group A data, Cluster 2 contained data from all three groups, and Cluster 3 contained data from only groups 1 and 2. For the categorically normalized data sets there was a more uniform distribution of group data in the three Clusters. A less pronounced effect was seen in the rescaled image data features. Conclusion: Image Rescaling and feature renormalization can have a significant effect on the results of clustering analysis. These effects are also likely to influence the results of supervised machine learning algorithms. It may be possible to partly remove the influence of scanner field strength and the presence of Gadolinium based contrast in feature extraction for radiomics applications.

60 APPLIED LIFE SCIENCES↗

SU-F-R-52: A Comparison of the Performance of Radiomic Features From Free Breathing and 4DCT Scans in Predicting Disease Recurrence in Lung Cancer SBRT Patients

Purpose: There is a clinical need to identify patients who are at highest risk of recurrence after being treated with stereotactic body radiation therapy (SBRT). Radiomics offers a non-invasive approach by extracting quantitative features from medical images based on tumor phenotype that is predictive of an outcome. Lung cancer patients treated with SBRT routinely undergo free breathing (FB image) and 4DCT (average intensity projection (AIP) image) scans for treatment planning to account for organ motion. The aim of the current study is to evaluate and compare the prognostic performance of radiomic features extracted from FB and AIP images in lung cancer patients treated with SBRT to identify which image type would generate an optimal predictive model for recurrence. Methods: FB and AIP images of 113 Stage I-II NSCLC patients treated with SBRT were analysed. The prognostic performance of radiomic features for distant metastasis (DM) was evaluated by their concordance index (CI). Radiomic features were compared with conventional imaging metrics (e.g. diameter). All p-values were corrected for multiple testing using the false discovery rate. Results: All patients received SBRT and 20.4% of patients developed DM. From each image type (FB or AIP), nineteen radiomic features were selected based on stability and variance. Both image types had five common and fourteen different radiomic features. One FB (CI=0.70) and five AIP (CI range=0.65–0.68) radiomic features were significantly prognostic for DM (p<0.05). None of the conventional features derived from FB images (range CI=0.60–0.61) were significant but all AIP conventional features were (range CI=0.64–0.66). Conclusion: Features extracted from different types of CT scans have varying prognostic performances. AIP images contain more prognostic radiomic features for DM than FB images. These methods can provide personalized medicine approaches at low cost, as FB and AIP data are readily available within a large number of radiation oncology departments. R.M. had consulting interest with Amgen (ended in 2015).

60 APPLIED LIFE SCIENCES↗

Stimulus information contaminates summation tests of independent neural representations of features

Many models of visual processing assume that visual information is analyzed into separable and independent neural codes, or features. A common psychophysical test of independent features is known as a summation study, which measures performance in a detection, discrimination, or visual search task as the number of proposed features increases. Improvement in human performance with increasing number of available features is typically attributed to the summation, or combination, of information across independent neural coding of the features. In many instances, however, increasing the number of available features also increases the stimulus information in the task, as assessed by an optimal observer that does not include the independent neural codes. In a visual search task with spatial frequency and orientation as the component features, a particular set of stimuli were chosen so that all searches had equivalent stimulus information, regardless of the number of features. In this case, human performance did not improve with increasing number of features, implying that the improvement observed with additional features may be due to stimulus information and not the combination across independent features.

NASA Discipline Space Human Factors↗

Confidence-Based Feature Acquisition

Confidence-based Feature Acquisition (CFA) is a novel, supervised learning method for acquiring missing feature values when there is missing data at both training (learning) and test (deployment) time. To train a machine learning classifier, data is encoded with a series of input features describing each item. In some applications, the training data may have missing values for some of the features, which can be acquired at a given cost. A relevant JPL example is that of the Mars rover exploration in which the features are obtained from a variety of different instruments, with different power consumption and integration time costs. The challenge is to decide which features will lead to increased classification performance and are therefore worth acquiring (paying the cost). To solve this problem, CFA, which is made up of two algorithms (CFA-train and CFA-predict), has been designed to greedily minimize total acquisition cost (during training and testing) while aiming for a specific accuracy level (specified as a confidence threshold). With this method, it is assumed that there is a nonempty subset of features that are free; that is, every instance in the data set includes these features initially for zero cost. It is also assumed that the feature acquisition (FA) cost associated with each feature is known in advance, and that the FA cost for a given feature is the same for all instances. Finally, CFA requires that the base-level classifiers produce not only a classification, but also a confidence (or posterior probability).

Wagstaff, Kiri L.↗

Performance evaluation of automated data-driven feature extraction and selection methods for practical and scalable building energy consumption prediction models

Here, this study quantifies the impact of automated feature engineering methods (feature extraction and selection) on the quality and accuracy of machine learning models that predict building energy consumption. The case study compares model performance for three main scenarios: baseline (no feature extraction and selection), feature extraction only, and feature extraction combined with feature selection (filter and/or wrapper methods) for fully trained machine learning models for 200 metered/sub-metered energy measurements across 118 real buildings. For consistency, the same machine learning model architecture (a black box deep learning neural network with probabilistic forecast output) was used for all scenarios. Based on results, all feature engineering methods provided noticeable prediction accuracy improvements (e.g., 29%-68% median prediction improvement) compared to baseline scenarios. However, in this application, feature selection methods provide little practical value due to their limited performance gains and high computational cost. Smarter algorithm development supported by better computational environments will be needed before feature selection methods can reliably and efficiently improve predictive model performance.

97 MATHEMATICS AND COMPUTING↗

Spatial-Temporal [{sup 18}F]FDG-PET Features for Predicting Pathologic Response of Esophageal Cancer to Neoadjuvant Chemoradiation Therapy

Purpose: To extract and study comprehensive spatial-temporal {sup 18}F-labeled fluorodeoxyglucose ([{sup 18}F]FDG) positron emission tomography (PET) features for the prediction of pathologic tumor response to neoadjuvant chemoradiation therapy (CRT) in esophageal cancer. Methods and Materials: Twenty patients with esophageal cancer were treated with trimodal therapy (CRT plus surgery) and underwent [{sup 18}F]FDG-PET/CT scans both before (pre-CRT) and after (post-CRT) CRT. The 2 scans were rigidly registered. A tumor volume was semiautomatically delineated using a threshold standardized uptake value (SUV) of ≥2.5, followed by manual editing. Comprehensive features were extracted to characterize SUV intensity distribution, spatial patterns (texture), tumor geometry, and associated changes resulting from CRT. The usefulness of each feature in predicting pathologic tumor response to CRT was evaluated using the area under the receiver operating characteristic curve (AUC) value. Results: The best traditional response measure was decline in maximum SUV (SUV{sub max}; AUC, 0.76). Two new intensity features, decline in mean SUV (SUV{sub mean}) and skewness, and 3 texture features (inertia, correlation, and cluster prominence) were found to be significant predictors with AUC values ≥0.76. According to these features, a tumor was more likely to be a responder when the SUV{sub mean} decline was larger, when there were relatively fewer voxels with higher SUV values pre-CRT, or when [{sup 18}F]FDG uptake post-CRT was relatively homogeneous. All of the most accurate predictive features were extracted from the entire tumor rather than from the most active part of the tumor. For SUV intensity features and tumor size features, changes were more predictive than pre- or post-CRT assessment alone. Conclusion: Spatial-temporal [{sup 18}F]FDG-PET features were found to be useful predictors of pathologic tumor response to neoadjuvant CRT in esophageal cancer.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Lung Texture in Serial Thoracic Computed Tomography Scans: Correlation of Radiomics-based Features With Radiation Therapy Dose and Radiation Pneumonitis Development

Purpose: To assess the relationship between radiation dose and change in a set of mathematical intensity- and texture-based features and to determine the ability of texture analysis to identify patients who develop radiation pneumonitis (RP). Methods and Materials: A total of 106 patients who received radiation therapy (RT) for esophageal cancer were retrospectively identified under institutional review board approval. For each patient, diagnostic computed tomography (CT) scans were acquired before (0-168 days) and after (5-120 days) RT, and a treatment planning CT scan with an associated dose map was obtained. 32- × 32-pixel regions of interest (ROIs) were randomly identified in the lungs of each pre-RT scan. ROIs were subsequently mapped to the post-RT scan and the planning scan dose map by using deformable image registration. The changes in 20 feature values (ΔFV) between pre- and post-RT scan ROIs were calculated. Regression modeling and analysis of variance were used to test the relationships between ΔFV, mean ROI dose, and development of grade ≥2 RP. Area under the receiver operating characteristic curve (AUC) was calculated to determine each feature's ability to distinguish between patients with and those without RP. A classifier was constructed to determine whether 2- or 3-feature combinations could improve RP distinction. Results: For all 20 features, a significant ΔFV was observed with increasing radiation dose. Twelve features changed significantly for patients with RP. Individual texture features could discriminate between patients with and those without RP with moderate performance (AUCs from 0.49 to 0.78). Using multiple features in a classifier, AUC increased significantly (0.59-0.84). Conclusions: A relationship between dose and change in a set of image-based features was observed. For 12 features, ΔFV was significantly related to RP development. This study demonstrated the ability of radiomics to provide a quantitative, individualized measurement of patient lung tissue reaction to RT and assess RP development.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Hybrid feature selection approach for power transformer fault diagnosis based on support vector machine and genetic algorithm

To further improve fault diagnosis accuracy, a new hybrid feature selection approach combined with a genetic algorithm (GA) and support vector machine (SVM) is presented in this study. Adaptive synthetic technique and arctangent transformation method are adopted to improve the statistical property of the training set (IEC TC10 dataset). Five filter methods based on different evaluation metrics are employed to rank 48 input features derived from dissolved gas analysis (DGA). Then, feature combination methods are applied to aggregate feature ranks and form a lower-dimension candidate feature subset. The GA-SVM model is implemented to optimise parameters and select optimal feature subsets. 5-fold cross-validation accuracy of the GA-SVM is used to evaluate fault diagnosis capability of feature subsets and finally, a novel subset is determined as the optimal feature subset. Accuracy comparison manifests the superiority of the optimal feature subsets over that of conventional approaches. Besides, generalisation and robustness of the optimal subset are validated by testing DGA samples from the local power utility. Results indicate that the optimal feature subset obtained by the proposed method can significantly improve the accuracies of power transformer fault diagnosis.

Kari, Tusongjiang↗

TU-A-12A-10: Robust Radiomics Feature Quantification Using Semiautomatic Volumetric Segmentation

Purpose: Recent advances in medical imaging technologies provide opportunities to quantify the tumor phenotype throughout the course of treatment non-invasively. The emerging field of Radiomics addresses this by converting medical images into minable data by applying a large number of quantitative imaging algorithms. Accurate tumor segmentation is one of the main challenges of Radiomics. It has been shown that semiautomatic segmentation approaches efficiently reduce inter-observer variability as compared to the time consuming manual delineations. In this study, a semiautomatic volumetric segmentation algorithm, implemented in the free and publicly available 3D-Slicer platform, was investigated in terms of its robustness for Radiomics features quantification. Methods: Fifty-six 3D-Radiomics features, quantifying phenotypic differences based on the tumor intensity, shape and texture, were extracted from the computed tomography images of twenty lung cancer patients. These Radiomics features were derived from the 3D-tumor volumes defined by three independent observers twice using 3D-Slicer, and compared to manual slice-by-slice delineations of five independent physicians in terms of intra-class correlation coefficient (ICC) and feature range. Results: Radiomics features extracted from 3D-Slicer segmentations had significantly higher reproducibility (ICC= 0.85 0.15, p=0.0009) compared to the features extracted from the manual segmentations (ICC= 0.77 0.17). Furthermore, the features extracted from 3D-Slicer based segmentations, spread over significantly smaller range across observers as compared to those of the manual delineations (p= 3.819e-07). Moreover, the features derived from 3D-Slicer segmentations overlapped in range with those of the manual delineations, as the lower(higher) limit(s) being significantly higher(lower) for the 3D-Slicer features (p= 0.007, p= 5.863e-06). Conclusion: 3D-Slicer based semiautomatic segmentation significantly improves the robustness of radiomics features and thus could serve as a potential alternative to the time consuming manual segmentation process. So 3D-Slicer can have a large application in Radiomics, and be employed for high-throughput data mining research of medical imaging in clinical oncology.

60 APPLIED LIFE SCIENCES↗

SU-E-J-260: Quantitative Image Feature Analysis of Multiphase Liver CT for Hepatocellular Carcinoma (HCC) in Radiation Therapy

Purpose: To identify the effective quantitative image features (radiomics features) for prediction of response, survival, recurrence and metastasis of hepatocellular carcinoma (HCC) in radiotherapy. Methods: Multiphase contrast enhanced liver CT images were acquired in 16 patients with HCC on pre and post radiation therapy (RT). In this study, arterial phase CT images were selected to analyze the effectiveness of image features for the prediction of treatment outcome of HCC to RT. Response evaluated by RECIST criteria, survival, local recurrence (LR), distant metastasis (DM) and liver metastasis (LM) were examined. A radiation oncologist manually delineated the tumor and normal liver on pre and post CT scans, respectively. Quantitative image features were extracted to characterize the intensity distribution (n=8), spatial patterns (texture, n=36), and shape (n=16) of the tumor and liver, respectively. Moreover, differences between pre and post image features were calculated (n=120). A total of 360 features were extracted and then analyzed by unpaired student’s t-test to rank the effectiveness of features for the prediction of response. Results: The five most effective features were selected for prediction of each outcome. Significant predictors for tumor response and survival are changes in tumor shape (Second Major Axes Length, p= 0.002; Eccentricity, p=0.0002), for LR, liver texture (Standard Deviation (SD) of High Grey Level Run Emphasis and SD of Entropy, both p=0.005) on pre and post CT images, for DM, tumor texture (SD of Entropy, p=0.01) on pre CT image and for LM, liver (Mean of Cluster Shade, p=0.004) and tumor texture (SD of Entropy, p=0.006) on pre CT image. Intensity distribution features were not significant (p>0.09). Conclusion: Quantitative CT image features were found to be potential predictors of the five endpoints of HCC in RT. This work was supported in part by the National Cancer Institute Grant R01CA172638.

60 APPLIED LIFE SCIENCES↗

SU-E-J-264: Using Magnetic Resonance Imaging-Derived Features to Quantify Radiotherapy-Induced Normal Tissue Morbidity

Purpose: The aim of this study was to explore the use of Magnetic Resonance Imaging (MRI)-derived features as indicators of Radiotherapy (RT)-induced normal tissue morbidity. We also investigate the relationship between these features and RT dose in four critical structures. Methods: We demonstrate our approach for four patients treated with RT for base of tongue cancer in 2005–2007. For each patient, two MRI scans (T1-weighted pre (T1pre) and post (T1post) gadolinium contrast-enhancement) were acquired within the first six months after RT. The assessed morbidity endpoint observed in 2/4 patients was Grade 2+ CTCAEv.3 trismus. Four ipsilateral masticatory-related structures (masseter, lateral and medial pterygoid, and the temporal muscles) were delineated on both T1pre and T1post and these scans were co-registered to the treatment planning CT using a deformable demons algorithm. For each structure, the maximum and mean RT dose, and six MRI-derived features (the second order texture features entropy and homogeneity, and the first order mean, median, kurtosis, and skewness) were extracted and compared structure-wise between patients with and without trismus. All MRI-derived features were calculated as the difference between T1pre and T1post, ΔS. Results: For 5/6 features and all structures, ΔS diverged between trismus and non-trismus patients particularly for the masseter, lateral pterygoid, and temporal muscles using the kurtosis feature (−0.2 vs. 6.4 for lateral pterygoid). Both the maximum and mean RT dose in all four muscles were higher amongst the trismus patients (with the maximum dose being up to 25 Gy higher). Conclusion: Using MRI-derived features to quantify RT-induced normal tissue complications is feasible. We showed that several features are different between patients with and without morbidity and that the RT dose in all investigated structures are higher amongst patients with morbidity. MRI-derived features, therefore, has the potential to improve predictions of normal tissue morbidity.

60 APPLIED LIFE SCIENCES↗

TU-CD-BRB-01: Normal Lung CT Texture Features Improve Predictive Models for Radiation Pneumonitis

Purpose: Existing normal tissue complication probability (NTCP) models for radiation pneumonitis (RP) traditionally rely on dosimetric and clinical data but are limited in terms of performance and generalizability. Extraction of pre-treatment image features provides a potential new category of data that can improve NTCP models for RP. We consider quantitative measures of total lung CT intensity and texture in a framework for prediction of RP. Methods: Available clinical and dosimetric data was collected for 198 NSCLC patients treated with definitive radiotherapy. Intensity- and texture-based image features were extracted from the T50 phase of the 4D-CT acquired for treatment planning. A total of 3888 features (15 clinical, 175 dosimetric, and 3698 image features) were gathered and considered candidate predictors for modeling of RP grade≥3. A baseline logistic regression model with mean lung dose (MLD) was first considered. Additionally, a least absolute shrinkage and selection operator (LASSO) logistic regression was applied to the set of clinical and dosimetric features, and subsequently to the full set of clinical, dosimetric, and image features. Model performance was assessed by comparing area under the curve (AUC). Results: A simple logistic fit of MLD was an inadequate model of the data (AUC∼0.5). Including clinical and dosimetric parameters within the framework of the LASSO resulted in improved performance (AUC=0.648). Analysis of the full cohort of clinical, dosimetric, and image features provided further and significant improvement in model performance (AUC=0.727). Conclusions: To achieve significant gains in predictive modeling of RP, new categories of data should be considered in addition to clinical and dosimetric features. We have successfully incorporated CT image features into a framework for modeling RP and have demonstrated improved predictive performance. Validation and further investigation of CT image features in the context of RP NTCP modeling is warranted. This work was supported by the Rosalie B. Hite Fellowship in Cancer research awarded to SPK.

60 APPLIED LIFE SCIENCES↗