Home » Key Medical Diagnostics Articles » Test-Retest Reproducibility Analysis of Lung CT Image Features

Test-Retest Reproducibility Analysis of Lung CT Image Features

Significance Statement

Quantitative size, shape and texture features derived from computed tomographic (CT) images may be useful as predictive, prognostic or response biomarkers in non-small cell lung cancer (NSCLC). However, to be useful, such features must be reproducible, non-redundant and have a large dynamic range.  We developed a set of quantitative three-dimensional (3D) features to describe segmented tumors and evaluated their reproducibility to select features with high potential to have prognostic utility.   In this study thirty two patients with unenhanced CT scans of NSCLC patients with primary lung cancer lesions were segmented using semi-automatic 3-D region growing algorithms. Following segmentation, 219 quantitative 3D features were extracted from each lesion, corresponding to size, shape, and texture, including features in transformed spaces (Laws, wavelets). The most informative features were selected using the concordance correlation coefficient across test-retest, the biological range and a feature independence measure.    In this study [1] as illustrated in figure 1, we first showed a procedure to obtained informative image features and then use those to predict radiological prognosis. Of the 66 concordant features, 42 features were non-redundant after grouping features with R2Bet ≥ 0.95. These reproducible features were found to be predictive of radiological prognosis. The area under the curve (AUC) was 91% for a size based feature and 92% for the texture features (Runlength, Laws). We tested the ability of image features to predict a radiological prognostic score on an independent NSCLC (39 adenocarcinoma)  samples (see figure 2), the AUC for texture features (Runlength Emphasis, Energy) was 0.84 while the conventional size based features (Volume, longest diameter) was 0.80.  In the follow up study [2], we compared manual (expert driven) and ensemble or semi-automated (seed point) delineation methods to find reproducible and non-redundant quantitative image features and tested for prognostic prediction.  At a stringent setting, there were 29 features across segmentation methods found to be repeatable and non-redundant at higher cutoff (CCCTreT & DR ≥ 0.9 & R2Bet ≥ 0.95).  The representative features were tested for their prognostic capabilities in an independent non-small cell lung cancer data set (59 Lung Adenocarcinomas), where one of the texture features, Runlength GLN was statistically significant in separating the samples into survival groups (p-value of 0.046). Figure 3 shows samples in the two prognostic groups.

 

[1]  Y Balagurunathan, V Kumar, Y Gu, J Kim, H Wang, Y Liu, D Goldgof, L Hall, R Korn, B Zhao, L Schwartz, S Basu, S Eschrich, R Gatenby, R Gillies. Test Retest reproducibility analysis of lung CT image features. Journal of Digital Imaging, 27, 805-823, [Epub. Jun 2014] Dec 2014.

[2] Y Balagurunathan, Y Gu, H Wang, V Kumar, O Grove, S. Hawkins, J. Kim, D. Goldgof, L. Hall, R. Gatenby, R. Gillies, Reproducibility and Prognosis of quantitative features extracted from CT images of lung tumor, J. Trans. Oncol.,7(1), 72-87,  Feb 2014.

 

correspondence: {Robert.gillies, yogab}@moffitt.org

Test-Retest Reproducibility Analysis of Lung CT Image Feature11111111

 

 

 

 

 

 

 

 

Figure 1. Process flow in finding reproducible image features that are predictive of radiological prognosis.

Test-Retest Reproducibility Analysis of Lung CT Image Features22222

 

 

 

 

 

 

Figure 2. Three representative slices of patient CT scans selected from 39 adenocarcinoma  cases with better radiological prognostic score (A&B, score of 0.44 & 0.48 respectively) and poor radiological prognostic score (C &D, score of 0.8, 0.92 respectively).

Test-Retest Reproducibility Analysis of Lung CT Image Features

 

 

 

Figure 3. Prognostic test result using Run length Non-uniformity (GLN) image feature split at the median value is shown. Top row correspond to better prognosis compared to bottom row.

 

Journal Reference

Balagurunathan Y, Kumar V, Gu Y, Kim J, Wang H, Liu Y, Goldgof DB, Hall LO, Korn R, Zhao B, Schwartz LH, Basu S, Eschrich S, Gatenby RA, Gillies RJ.

J Digit Imaging. 2014 Dec;27(6):805-23.

Department of Cancer Imaging and Metabolism, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA.

 

Abstract

Quantitative size, shape, and texture features derived from computed tomographic (CT) images may be useful as predictive, prognostic, or response biomarkers in non-small cell lung cancer (NSCLC). However, to be useful, such features must be reproducible, non-redundant, and have a large dynamic range. We developed a set of quantitative three-dimensional (3D) features to describe segmented tumors and evaluated their reproducibilityto select features with high potential to have prognostic utility. Thirty-two patients with NSCLC were subjected to unenhanced thoracic CT scans acquired within 15 min of each other under an approved protocol. Primary lung cancer lesions were segmented using semi-automatic 3D region growing algorithms. Following segmentation, 219 quantitative 3D features were extracted from each lesion, corresponding to size, shape, and texture, including features in transformed spaces (laws, wavelets). The most informative features were selected using the concordance correlation coefficient across test-retest, the biological range and a feature independence measure. There were 66 (30.14 %) features with concordance correlation coefficient ≥ 0.90 across test-retest and acceptable dynamic range. Of these, 42 features were non-redundant after grouping features with R (2) Bet ≥ 0.95. These reproducible features were found to be predictive of radiological prognosis. The area under the curve (AUC) was 91 % for a size-based feature and 92 % for the texture features (runlength, laws). We tested the ability of image features to predict a radiological prognostic score on an independent NSCLC (39 adenocarcinoma) samples, the AUC for texture features (runlength emphasis, energy) was 0.84 while the conventional size-based features (volume, longest diameter) was 0.80. Test-retest and correlation analyses have identified non-redundant  CT image features with both high intra-patient reproducibility and inter-patient biological range. Thus making the case that quantitative  image  features are informative and prognostic biomarkers for NSCLC.

Go To PubMed