Automatic classification of centrilobular emphysema on CT using deep learning: Comparison with visual scoring

Abstract

The presence and severity of emphysema, scored visually on computed tomography (CT) using a classification system developed by the Fleischner Society, is a clinically significant index of disease severity. Since visual assessment can be subjective and is time consuming, our purpose was to evaluate the potential of a deep learning method for automatic grading of emphysema. The study cohort included 8213 subjects enrolled in the COPDGene study. Baseline CT and visual scores on 2500 subjects were used to train a deep learning model for classification of centrilobular emphysema according to the Fleischner system. The model was then used to predict emphysema scores on 5713 subjects not included in the training set. Predictions were compared with visual emphysema scores, pulmonary function tests (PFTs), smoking history and St. George Respiratory Questionnaire (SGRQ). Agreement between visual emphysema scores and those generated automatically was moderate (weighted κ = 0.60, p < 0.0001). Emphysema scores predicted by the deep learning model showed significant associations with PFTs, smoking history and SGRQ, similar to those seen in comparison with visual scores.