Texture-based Automated Quantification of Interstitial Lung Disease: Correlation With the Visual Score

Purpose of the Study: The quantification of various disease patterns in the lung parenchyma remains a challenge. In this study the texture analysis algorithm 3D-AMFM (Adaptive Multiple Feature Method) contained in the software PASS (University of Iowa) was applied with interstitial lung disease (ILD). We checked for the statistical accuracy and reliabity of the method by standard tests compared to visual scoring. Methods: Based on a Bayesian classifier, a training data base including texture patterns (normal, ground glass, honey combing, emphysema) from 1300 volumes of interest (VOIs; 151515 pixels) of 47 selected patients with mixed ILDs was built up. Another 18 patients with a typical thin-section CT pattern of usual interstitial pneumonia (UIP) (n=9) and nonspecific interstitial pneumonia (NSIP) (n=9) were independently analyzed and visually quantified at 5 pre-established levels by two experienced chest radiologists. The same thin-section CT scans were analyzed with 3D-AMFM. Wilcoxon test was used to evaluate the correlation between the visual scores and the computed results. Results: The mean extent of honeycombing, ground glass and emphysema was 5.4%, 43.5% and 2.1% by the visual score and 19.4%, 44.3% and 0.6% by the 3D-AMFM, respectively. There was close correlation between visual score and 3D-AMFM for both the extent of ground glass (P=0.546) and emphysema (P=0.099), but worse for the extent of honey combing (P=0.000837). Conclusions: The 3D-AMFM system is a promising and effective tool for ILD quantification, showing clinical acceptable correlation with human observer. The overestimation of honeycombing by 3D-AMFM is probably caused by small vessels and airways. The continuing development of the feature data base and the inclusion of further pathologic texture patterns will improve quantification of disease and provide objective measures of disease progression.