Interstitial lung abnormalities (ILA) on routine chest CT: Comparison of radiologists' visual evaluation and automated quantification

Abstract

Purpose: We aimed to evaluate the performance of a fully automated quantitative software in detecting interstitial lung abnormalities (ILA) according to the Fleischner Society guidelines on routine chest CT compared with radiologists' visual analysis.

Method: This retrospective single-centre study included participants with ILA findings and 1:2 matched controls who underwent routine chest CT using various CT protocols for health screening. Two thoracic radiologists independently reviewed the CT images using the Fleischner Society guidelines. We developed a fully automated quantitative tool for detecting ILA by modifying deep learning-based quantification of interstitial lung disease and evaluated its performance using the radiologists' consensus for ILA as a reference standard.

Results: A total of 336 participants (mean age, 70.5 ± 6.1 years; M:F = 282:54) were included. Inter-reader agreements were substantial for the presence of ILA (weighted κ, 0.74) and fair for its subtypes (weighted κ, 0.38). The quantification system for identifying ILA using a threshold of 5 % in at least one zone showed 67.6 % sensitivity, 93.3 % specificity, and 90.5 % accuracy. Eight of 20 (40 %) false positives identified by the system were underestimated by readers for ILA extent. Contrast-enhancement in a certain vendor and suboptimal inspiration caused a true false-positive on the system (all P < 0.05). The best cut-off value of abnormality extent detecting ILA on the system was 3.6 % (sensitivity, 84.8 %; specificity 92.4 %).

Conclusions: Inter-reader agreement was substantial for ILA but only fair for its subtypes. Applying an automated quantification system in routine clinical practice may aid the objective identification of ILA.

Keywords: Deep learning; Interstitial lung abnormality; Interstitial lung disease; Quantification.