Computed Tomography–Based Deep Learning Model for Assessing the Severity of Patients With Connective Tissue Disease–Associated Interstitial Lung Disease

Abstract

Objectives This study aimed to develop a computed tomography (CT)–based deep learning model for assessing the severity of patients with connective tissue disease (CTD)–associated interstitial lung disease (ILD). Methods The retrospective study included 298 CTD-ILD patients between January 2018 and May 2022. A deep learning–based RDNet model was established (1610 fully annotated CT images for training and 402 images for validation). The model was used to automatically classify and quantify 3 radiologic features (ground glass opacities [GGOs], reticulation, and honeycombing), along with a volumetric sum of 3 areas (ILD%). As a control, we used 4 previously defined CT threshold methods to calculate the ILD assessment index. The Spearman rank correlation coefficient (r) evaluated the correlation between various indicators and the lung function index in the remaining 184 CTD-ILD patients who were staged according to the gender-age-physiology (GAP) system. Results The RDNet model accurately identified GGOs, reticulation, and honeycombing, with corresponding Dice indexes of 0.784, 0.782, and 0.747, respectively. A total of 137 patients were at GAP1 (73.9%), 36 patients at GAP2 (19.6%), and 11 patients at GAP3 (6.0%). The percentages of reticulation and honeycombing at GAP2 and GAP3 were markedly elevated compared with those at GAP1 (P < 0.001). The percentage of GGOs was not significantly different among the GAP stages (P = 0.62). As the GAP stage increased, all lung function indicators tended to decrease, and the composite physiologic index (CPI) indicated an upward tendency. The percentage of honeycombs moderately correlated with the percentage of diffusing capacity of the lung for carbon monoxide (DLco%) (r = −0.58, P < 0.001) and CPI (r = 0.63, P < 0.001). The ILD assessment index calculated by the CT threshold method (−260 to −600 Hounsfield units) had a low correlation with DLco% and CPI (DLco%: r = −0.42, P < 0.001; CPI: r = 0.45, P < 0.001). Conclusions The RDNet model can quantify GGOs, reticulation, and honeycombing of chest CT images in CTD-ILD patients, among which honeycombing had the most significant effect on lung function indicators. In addition, this model provided good clinical utility for evaluating the severity of CTD-ILD.