Neighborhood Gray-tone Difference Matrix Research Articles

Impact of image preprocessing on the volume dependence and prognostic potential of radiomics features in non-small cell lung cancer

Background: Radiomics features have been used in a variety of studies to predict patient outcomes or aid in the diagnosis of non-small cell lung cancer. However, no guidelines exist for the best way to calculate these features to maximize their prognostic potential. The purpose of the current study was to evaluate how different image pre-processing techniques may impact both the volume dependence and prognostic potential of the features in univariate analyses. Methods: Radiomics features from the histogram, co-occurrence matrix, neighborhood gray-tone difference matrix and run-length matrix were calculated from a set of computed tomography (CT) images of 107 non-small cell lung cancer tumors with volumes ranging from 5 to 567 cm3. Features were calculated from the images with no pre-processing, 8 bit depth resampling, Butterworth smoothing, or both 8 bit depth resampling and Butterworth smoothing. To determine which features were correlated with volume, we calculated the Spearman rank correlation coefficient (rs) for each feature and preprocessing combination. For features that had very high volume correlations (rs >0.95) regardless of which preprocessing algorithm was used, we normalized the algorithm for volume and recalculated the volume correlation. To determine whether the preprocessing technique affected the usefulness of the feature, we fitted univariate Cox proportional hazards models for all four preprocessing techniques for each feature and calculated the P value. Additionally, univariate cox models were recalculated using leave-one-out cross validation to generate risk predictions for each patient. As a result, each patient had a predicted outcome for each model in which the patient was not involved in the model building. The prediction accuracy was assessed using Harrell’s concordance index (c-index). Finally, the ability of each feature to improve model fit was examined using the P value of the log-likelihood ratio between a model built using volume only and a model built using volume and one radiomics feature. The Benjamini-Hochberg procedure was used for multiplicity correction. Results: Five features were entirely volume dependent (busyness, coarseness, grey-level non-uniformity, run-length non-uniformity, and energy) and new algorithms were proposed for these features. Both the correlation with volume and the prognostic value of individual features changed substantially with different preprocessing techniques. In general, preprocessed features that were at least slightly correlated with volume (rs >0.5) were more likely to be significant in the univariate analysis. Additionally, Butterworth smoothing, used either alone or in conjunction with 8 bit depth resampling, most often yielded features that were significant in univariate analysis. Conclusions: Preprocessing can have a strong impact on the volume dependence of a feature, and its significance in univariate models. To create standardized features useful for multivariate modeling, it will be important to balance the usefulness of features with their volume dependence.

Texture based classification of the severity of mitral regurgitation

Clinically, the severity of valvular regurgitation is assessed by manual tracing of the regurgitant jet in the respective chambers. This work presents a computer-aided diagnostic (CAD) system for the assessment of the severity of mitral regurgitation (MR) based on image processing that does not require the intervention of the radiologist or clinician. Eight different texture feature sets from the regurgitant area (selected through an arbitrary criterion) have been used in the present approach. First order statistics have been used initially, however, observing their limitations, the other texture features such as spatial gray level difference matrix, gray level difference statistics, neighborhood gray tone difference matrix, statistical feature matrix, Laws’ textures energy measure, fractal dimension texture analysis and Fourier power spectrum have additionally been used. For the classification task a supervised classifier i.e., support vector machine has been used in the present approach. The classification accuracy has been improved significantly by using these texture features in combination, in comparison to when fed individually as input to the classifier. The classification accuracy of 95.65±1.09, 95.65±1.09 and 95.36±1.13 has been obtained in apical two chamber, apical four chamber and parasternal long axis views, respectively. Therefore, the results of this paper indicate that the proposed CAD system may effectively assist the radiologists in establishing (confirming) the MR stages, namely, mild, moderate and severe.

Differentiating the grades of thymic epithelial tumor malignancy using textural features of intratumoral heterogeneity via (18)F-FDG PET/CT.

We aimed to explore the ability of textural heterogeneity indices determined by (18)F-FDG PET/CT for grading the malignancy of thymic epithelial tumors (TETs). We retrospectively enrolled 47 patients with pathologically proven TETs who underwent pre-treatment (18)F-FDG PET/CT. TETs were classified by pathological results into three subgroups with increasing grades of malignancy: low-risk thymoma (LRT; WHO classification A, AB and B1), high-risk thymoma (B2 and B3), and thymic carcinoma (TC). Using (18)F-FDG PET/CT, we obtained conventional imaging indices including SUVmax and 20 intratumoral heterogeneity indices: i.e., four local-scale indices derived from the neighborhood gray-tone difference matrix (NGTDM), eight regional-scale indices from the gray-level run-length matrix (GLRLM), and eight regional-scale indices from the gray-level size zone matrix (GLSZM). Area under the receiver operating characteristic curve (AUC) was used to demonstrate the abilities of the imaging indices for differentiating subgroups. Multivariable logistic regression analysis was performed to show the independent significance of the textural indices. Combined criteria using optimal cutoff values of the SUVmax and a best-performing heterogeneity index were applied to investigate whether they improved differentiation between the subgroups. Most of the GLRLM and GLSZM indices and the SUVmax showed good or fair discrimination (AUC >0.7) with best performance for some of the GLRLM indices and the SUVmax, whereas the NGTDM indices showed relatively inferior performance. The discriminative ability of some of the GLSZM indices was independent from that of SUVmax in multivariate analysis. Combined use of the SUVmax and a GLSZM index improved positive predictive values for LRT and TC. Texture analysis of (18)F-FDG PET/CT scans has the potential to differentiate between TET tumor grades; regional-scale indices from GLRLM and GLSZM perform better than local-scale indices from the NGTDM. The SUVmax and heterogeneity indices may have complementary value in differentiating TET subgroups.

Uncertainty analysis of quantitative imaging features extracted from contrast-enhanced CT in lung tumors

PurposeTo assess the uncertainty of quantitative imaging features extracted from contrast-enhanced computed tomography (CT) scans of lung cancer patients in terms of the dependency on the time after contrast injection and the feature reproducibility between scans. MethodsEight patients underwent contrast-enhanced CT scans of lung tumors on two sessions 2–7 days apart. Each session included 6 CT scans of the same anatomy taken every 15s, starting 50s after contrast injection. Image features based on intensity histogram, co-occurrence matrix, neighborhood gray-tone difference matrix, run-length matrix, and geometric shape were extracted from the tumor for each scan. Spearman's correlation was used to examine the dependency of features on the time after contrast injection, with values over 0.50 considered time-dependent. Concordance correlation coefficients were calculated to examine the reproducibility of each feature between times of scans after contrast injection and between scanning sessions, with values greater than 0.90 considered reproducible. ResultsThe features were found to have little dependency on the time between the contrast injection and the CT scan. Most features were reproducible between times of scans after contrast injection and between scanning sessions. Some features were more reproducible when they were extracted from a CT scan performed at a longer time after contrast injection. ConclusionThe quantitative imaging features tested here are mostly reproducible and show little dependency on the time after contrast injection.

SU-D-BRA-05: Toward Understanding the Robustness of Radiomics Features in CT

Purpose: To gauge the impact of inter-scanner variability on radiomics features in computed tomography (CT). Methods: We compared the radiomics features calculated for 17 scans of the specially designed Credence Cartridge Radiomics (CCR) phantom with those calculated for 20 scans of non–small cell lung cancer (NSCLC) tumors. The scans were acquired at four medical centers using General Electric, Philips, Siemens, and Toshiba CT scanners. Each center used its own routine thoracic imaging protocol. To produce a large dynamic range of radiomics feature values, the CCR phantom has 10 cartridges comprising different materials. The features studied were derived from the neighborhood gray-tone difference matrix or image intensity histogram. To quantify the significance of the inter-scanner variability, we introduced the metric “feature noise”, which compares the ratio of inter-scanner variability and inter-patient variability in decibels, positive values indicating substantial noise. We performed hierarchical clustering based to look for dependence of the features on the scan acquisition parameters. Results: For 5 of the 10 features studied, the inter-scanner variability was larger than the inter-patient variability. Of the 10 materials in the phantom, shredded rubber seemed to produce feature values most similar to those of the NSCLC tumors. The feature busyness had the greatest feature noise (14.3 dB), whereas texture strength had the least (−14.6 dB). Hierarchical clustering indicated that the features depended in part on the scanner manufacturer, image slice thickness, and pixel size. Conclusion: The variability in the values of radiomics features calculated for CT images of a radiomics phantom can be substantial relative to the variability in the values of these features calculated for CT images of NSCLC tumors. These inter-scanner differences and their effects should be carefully considered in future radiomics studies.

The precision of textural analysis in (18)F-FDG-PET scans of oesophageal cancer.

Measuring tumour heterogeneity by textural analysis in (18)F-fluorodeoxyglucose positron emission tomography ((18)F-FDG PET) provides predictive and prognostic information but technical aspects of image processing can influence parameter measurements. We therefore tested effects of image smoothing, segmentation and quantisation on the precision of heterogeneity measurements. Sixty-four (18)F-FDG PET/CT images of oesophageal cancer were processed using different Gaussian smoothing levels (2.0, 2.5, 3.0, 3.5, 4.0 mm), maximum standardised uptake value (SUVmax) segmentation thresholds (45%, 50%, 55%, 60%) and quantisation (8, 16, 32, 64, 128 bin widths). Heterogeneity parameters included grey-level co-occurrence matrix (GLCM), grey-level run length matrix (GLRL), neighbourhood grey-tone difference matrix (NGTDM), grey-level size zone matrix (GLSZM) and fractal analysis methods. The concordance correlation coefficient (CCC) for the three processing variables was calculated for each heterogeneity parameter. Most parameters showed poor agreement between different bin widths (CCC median 0.08, range 0.004-0.99). Segmentation and smoothing showed smaller effects on precision (segmentation: CCC median 0.82, range 0.33-0.97; smoothing: CCC median 0.99, range 0.58-0.99). Smoothing and segmentation have only a small effect on the precision of heterogeneity measurements in (18)F-FDG PET data. However, quantisation often has larger effects, highlighting a need for further evaluation and standardisation of parameters for multicentre studies. • Heterogeneity measurement precision in (18) F-FDG PET is influenced by image processing methods. • Quantisation shows large effects on precision of heterogeneity parameters in (18) F-FDG PET/CT. • Smoothing and segmentation show comparatively smaller effects on precision of heterogeneity parameters.

Surface Roughness Detection of Arteries via Texture Analysis of Ultrasound Images for Early Diagnosis of Atherosclerosis

There is a strong research interest in identifying the surface roughness of the carotid arterial inner wall via texture analysis for early diagnosis of atherosclerosis. The purpose of this study is to assess the efficacy of texture analysis methods for identifying arterial roughness in the early stage of atherosclerosis. Ultrasound images of common carotid arteries of 15 normal mice fed a normal diet and 28 apoE−/− mice fed a high-fat diet were recorded by a high-frequency ultrasound system (Vevo 2100, frequency: 40 MHz). Six different texture feature sets were extracted based on the following methods: first-order statistics, fractal dimension texture analysis, spatial gray level dependence matrix, gray level difference statistics, the neighborhood gray tone difference matrix, and the statistical feature matrix. Statistical analysis indicates that 11 of 19 texture features can be used to distinguish between normal and abnormal groups (p<0.05). When the 11 optimal features were used as inputs to a support vector machine classifier, we achieved over 89% accuracy, 87% sensitivity and 93% specificity. The accuracy, sensitivity and specificity for the k-nearest neighbor classifier were 73%, 75% and 70%, respectively. The results show that it is feasible to identify arterial surface roughness based on texture features extracted from ultrasound images of the carotid arterial wall. This method is shown to be useful for early detection and diagnosis of atherosclerosis.

Textural Features of Pretreatment 18F-FDG PET/CT Images: Prognostic Significance in Patients with Advanced T-Stage Oropharyngeal Squamous Cell Carcinoma

Previous studies have shown that total lesion glycolysis (TLG) may serve as a prognostic indicator in oropharyngeal squamous cell carcinoma (OPSCC). We sought to investigate whether the textural features of pretreatment (18)F-FDG PET/CT images can provide any additional prognostic information over TLG and clinical staging in patients with advanced T-stage OPSCC. We retrospectively analyzed the pretreatment (18)F-FDG PET/CT images of 70 patients with advanced T-stage OPSCC who had completed concurrent chemoradiotherapy, bioradiotherapy, or radiotherapy with curative intent. All of the patients had data on human papillomavirus (HPV) infection and were followed up for at least 24 mo or until death. A standardized uptake value (SUV) of 2.5 was taken as a cutoff for tumor boundary. The textural features of pretreatment (18)F-FDG PET/CT images were extracted from histogram analysis (SUV variance and SUV entropy), normalized gray-level cooccurrence matrix (uniformity, entropy, dissimilarity, contrast, homogeneity, inverse different moment, and correlation), and neighborhood gray-tone difference matrix (coarseness, contrast, busyness, complexity, and strength). Receiver-operating-characteristic curves were used to identify the optimal cutoff values for the textural features and TLG. Thirteen patients were HPV-positive. Multivariate Cox regression analysis showed that age, tumor TLG, and uniformity were independently associated with progression-free survival (PFS) and disease-specific survival (DSS). TLG, uniformity, and HPV positivity were significantly associated with overall survival (OS). A prognostic scoring system based on TLG and uniformity was derived. Patients who presented with TLG > 121.9 g and uniformity ≤ 0.138 experienced significantly worse PFS, DSS, and OS rates than those without (P < 0.001, < 0.001, and 0.002, respectively). Patients with TLG > 121.9 g or uniformity ≤ 0.138 were further divided according to age, and different PFS and DSS were observed. Uniformity extracted from the normalized gray-level cooccurrence matrix represents an independent prognostic predictor in patients with advanced T-stage OPSCC. A scoring system was developed and may serve as a risk-stratification strategy for guiding therapy.

P5-08-02: Mammographic Breast Texture Predicts Benign Biopsy Results and Composition.

Abstract Purpose: Investigate the relationship between pathology diagnosis, biopsy composition, and breast texture measured in digital mammogram in women who have had previous benign findings in their breast. The long-term goal of this project is to determine whether or not global breast texture is associated with tissue composition and biopsy results. Such a finding would be of interest because it would provide evidence that mammography images contain information related to breast biology beyond average density. Materials and Methods: A total of 124 women who had previously received benign biopsies underwent an additional dual-energy/low dose full-field digital mammography (DXA) scan of the unaffected breast. Three measures of breast density (percent dense area (PD), percent fibroglandular volume (%FGV), absolute fibroglandular volume (FGV) from mammogram, 15 biopsy types including apocrine metaplasia and non-secretory pituitary adenoma, and four tissue compositions (collagen percent area, fat percent area, ductal percent area, epithelial percent area) were estimated from biopsies for each participant. In addition 45 breast texture features were measured on each screening image taken during the dual-energy mammography using custom software. Tissue composition at the biopsy location was estimated based on the average percent area of stained collagen, fat, ductal, and epithelial from up to 3 slides taken from the biopsy. Pathology results from the three slides were taken to indicate the tissue type. Results: Neither mammographic percent density, %FGV, nor FGV showed significant association with either biopsy type or tissue composition. However, some of breast texture features were significantly associated with tissue type and tissue composition. The strongest associations were found between collagen percent area and the neighborhood gray tone difference matrix texture strength (R=0.26, P=0.005) and ductal percent area and the gray level contrast matrix homogeneity (GLCM HOM) (R=-0.30, P=0.001). When classified according to biopsy type, the univariate association between ductal percent area and GLCM HOM for apocrine metaplasia increased. (R=-0.55, P=0.007) Conclusion: 10 breast texture features of the unaffected breast were significantly associated with local tissue composition and biopsy type. Further study will be required to determine whether this association is related to masking effects related to breast complexity or changes in breast biology driving changes in breast morphology. CLINICAL RELEVANCE The use of texture measures as biomarkers of variations in breast biology is potentially clinically significant. This work indicates that texture measures are associated with some biological outcomes. It is possible that this work could be used to improve diagnostic utility of computer-aided diagnosis by weighting the probability of various benign outcomes according to the measured breast texture. Citation Information: Cancer Res 2011;71(24 Suppl):Abstract nr P5-08-02.

JAVA-based Texture Analysis Employing Neighborhood Gray-Tone Difference Matrix (NGTDM) for Optimization of Land Use Classifications in High Resolution Remote Sensing Data

JAVA-based Texture Analysis Employing Neighborhood Gray-Tone Difference Matrix (NGTDM) for Optimization of Land Use Classifications in High Resolution Remote Sensing Data

A Modular Software System to Assist Interpretation of Medical Images—Application to Vascular Ultrasound Images

Improvements in medical imaging technology have greatly contributed to early disease detection and diagnosis. However, the accuracy of an examination depends on both the quality of the images and the ability of the physician to interpret those images. Use of output from computerized analysis of an image may facilitate the diagnostic tasks and, potentially improve the overall interpretation of images and the subsequent patient care. In this paper, Analysis, a modular software system designed to assist interpretation of medical images, is described in detail. Analysis allows texture and motion estimation of selected regions of interest (ROIs). Texture features can be estimated using first-order statistics, second-order statistics, Laws' texture energy, neighborhood gray-tone difference matrix, gray level difference statistics, and the fractal dimension. Motion can be estimated from temporal image sequences using block matching or optical flow. Image preprocessing, manual and automatic definition of ROIs, and dimensionality reduction and clustering using fuzzy c-means, are also possible within Analysis. An important feature of Analysis is the possibility for online telecollaboration between health care professionals under a secure framework. To demonstrate the applicability and usefulness of the system in clinical practice, Analysis was applied to B-mode ultrasound images of the carotid artery. Diagnostic tasks included automatic segmentation of the arterial wall in transverse sections, selection of wall and plaque ROIs in longitudinal sections, estimation of texture features in different image areas, motion analysis of tissue ROIs, and clustering of the extracted features. It is concluded that Analysis can provide a useful platform for computerized analysis of medical images and support of diagnosis

Multifeature texture analysis for the classification of clouds in satellite imagery

The aim of this work was to develop a system based on multifeature texture analysis and modular neural networks that will facilitate the automated interpretation of satellite cloud images. Such a system will provide a standardized and efficient way for classifying cloud types that can be used as an operational tool in weather analysis. A series of 98 infrared satellite images from the geostationary satellite METEOSAT7 were employed, and 366 cloud segments were labeled into six cloud types after combined agreed observations from ground and satellite. From the segmented cloud images, nine different texture feature sets (a total of 55 features) were extracted, using the following algorithms: statistical features, spatial gray-level dependence matrices, gray-level difference statistics, neighborhood gray tone difference matrix, statistical feature matrix, Laws' texture energy measures, fractals, and Fourier power spectrum. The neural network self-organizing feature map (SOFM) classifier and the statistical K-nearest neighbor (KNN) classifier were used for the classification of the cloud images. Furthermore, the classification results of the nine different feature sets were combined, improving the classification yield for the six classes, for the SOFM classifier to 61% and for the KNN classifier to 64%.

Comparison of texture analysis methods for the characterization of coronary plaques in intravascular ultrasound images

Intravascular Ultrasound (IVUS) is a diagnostic imaging technique that provides tomographic visualization of coronary arteries. The aim of this study was to evaluate five texture analysis techniques and determine their ability to distinguish between plaque lesions of different composition. Using histological correlation, regions of calcified, fibrous, and necrotic core plaque were chosen from 27 coronary plaques. First-order statistics, Haralick's method, Laws' texture energy method, the neighborhood gray-tone difference matrix method, and texture spectrum features were examined using discriminant analysis. Self-validation indicated that Haralick's method yielded the most accurate results, with resubstitution and cross-validation error rates of 0.00 and 14.76%, respectively. Further optimization gave error rates of 6.67%, using only two discriminating features, IDM and entropy.