Automated detection of mucus plugs in CT imaging using a deep learning model

<bold>Background:</bold> Mucus plugs in obstructive airway diseases are associated with reduced pulmonary function, airway inflammation, and exacerbations. Qualitative assessment of mucus plug extent and severity may be evaluated using HRCT, but quantitative assessment remains time-consuming and observer-dependent. <bold>Aim:</bold> To evaluate the performance of a deep learning (DL) model trained to segment mucus plugs from chest CT. <bold>Methods:</bold> A 3D UNet DL segmentation model was trained on 287 manually annotated chest CTs from asthma patients. Plug count and mucus volume (mL) were derived from total intraluminal obstructions within the airway tree. The performance of the DL model was assessed in 19 scans from 10 COPD patients. Spearman correlations with FEV1%pred were computed using only the baseline scans. <bold>Results:</bold> A strong correlation was found between the DL model and the manual segmentations for both plug count (R=0.938) and volume (R=0.942). Considering the manual segmentations, FEV1%pred correlated significantly with the total plug count (ρ=-0.71, P=0.022) and total mucus volume (ρ=-0.70, P=0.024). Similarly, the deep learning model resulted in correlations between FEV1%pred and both total plug count ρ=-0.74, P=0.014) and total mucus volume ρ=-0.54, P=0.11). <bold>Conclusion:</bold> The deep learning model output was highly correlated with manual segmentations and correlated significantly with FEV1%pred. This proof-of-concept study shows that automation of this time-consuming process is viable and may expand its use in clinical research.