Automated quantification of high-resolution CT scan findings in individuals at risk for pulmonary fibrosis

Abstract

Background: Automated methods to quantify interstitial lung disease (ILD) on high-resolution CT (HRCT) scans in people at risk for pulmonary fibrosis have not been developed and validated.

Methods: Cohorts with familial pulmonary fibrosis (n = 126) or rheumatoid arthritis with and without ILD (n = 86) were used to develop and validate a computer program capable of quantifying ILD on HRCT scans, which imaged the lungs semicontinuously from the apices to the lung bases during end-inspiration in the prone position. This method uses segmentation, texture analysis, training, classification, and grading to score ILD.

Results: Quantification of HRCT scan findings of ILD using an automated computer program correlated with radiologist readings and detected disease of varying severity in a derivation cohort with familial pulmonary fibrosis or their first-degree relatives. This algorithm was validated in an independent cohort of subjects with rheumatoid arthritis with and without ILD. Automated classification of HRCT scans as normal or ILD was significant in the derivation and validation cohorts (P < .001 and P < .001, respectively). Areas under receiver operating characteristic curves performed independently for each group were 0.888 for the derivation cohort and 0.885 for the validation cohort. Pulmonary function test results, including FVC and diffusion capacity, correlated with computer-generated HRCT scan scores for ILD (r = -0.483 and r = -0.532, respectively).

Conclusions: Automated computer scoring of HRCT scans can objectively identify ILD and potentially quantify radiographic severity of lung disease in populations at risk for pulmonary fibrosis.

Figure 1.

Representative high-resolution CT (HRCT) scan images from four different subjects in the derivation cohort of individuals with familial pulmonary fibrosis or their relatives with or without familial interstitial lung disease (ILD). A, Minimal ILD. B, Mild ILD. C, Moderate ILD. D, Severe ILD. Arrows indicate representative areas of interstitial reticulation.

Figure 2.

A, Box-whisker plot of findings of ILD on HRCT scan of the chest in the derivation cohort reviewed by two independent radiologists compared with computer-generated scores. B, Box-whisker plot of computer scores and radiologist readings of ILD findings on HRCT scan of the chest in a validation cohort of subjects with rheumatoid arthritis with and without preclinical ILD or pulmonary fibrosis. The trends across groups (ie, no ILD, minimal, mild, moderate, severe) in both cohorts were significant. See Figure 1 legend for expansion of abbreviations.

Figure 3.

Receiver operating characteristic (ROC) curves plotting 1 minus specificity on the x-axis and sensitivity on the y-axis. A, Familial pulmonary fibrosis cohort. B, Rheumatoid arthritis cohort with and without ILD. C, Subjects in both cohorts. The area under the ROC curve was 0.888 for the familial pulmonary fibrosis cohort, 0.885 for the rheumatoid arthritis cohort with and without ILD, and 0.886 for both cohorts combined. The dashed identity line represents a CT score that performed no better than flipping a fair coin to classify CT scans as with or without ILD. See Figure 1 legend for expansion of abbreviation.

Figure 4.

A, Correlation of FVC % predicted with computer-generated scores in cohorts of subjects with familial pulmonary fibrosis and ILD associated with rheumatoid arthritis (r value = −0.483). B, Correlation of DLCO % predicted with computer-generated scores in cohorts of subjects with familial pulmonary fibrosis and ILD associated with rheumatoid arthritis (r value = −0.532). DLCO % = percentage of predicted diffusing capacity of the lung for carbon monoxide. See Figure 1 legend for expansion of other abbreviation.