Clinical Validation and Extension of an Automated, Deep Learning–Based Algorithm for Quantitative Sinus CT Analysis
Conner J. Massey, Laylaa Ramos, Daniel M. Beswick, Vijay R. Ramakrishnan, Stephen M. Humphries
Abstract
BACKGROUND AND PURPOSE: Sinus CT is critically important for the diagnosis of chronic rhinosinusitis. While CT is sensitive for detecting mucosal disease, automated methods for objective quantification of sinus opacification are lacking. We describe new measurements and further clinical validation of automated CT analysis using a convolutional neural network in a chronic rhinosinusitis population. This technology produces volumetric segmentations that permit calculation of percentage sinus opacification, mean Hounsfield units of opacities, and percentage of osteitis. MATERIALS AND METHODS: Demographic and clinical data were collected retrospectively from adult patients with chronic rhinosinusitis, including serum eosinophil count, Lund-Kennedy endoscopic scores, and the SinoNasal Outcomes Test-22. CT scans were scored using the Lund-Mackay score and the Global Osteitis Scoring Scale. CT images were automatically segmented and analyzed for percentage opacification, mean Hounsfield unit of opacities, and percentage osteitis. These readouts were correlated with visual scoring systems and with disease parameters using the Spearman ρ . RESULTS: Eighty-eight subjects were included. The algorithm successfully segmented 100% of scans and calculated features in a diverse population with CT images obtained on different scanners. A strong correlation existed between percentage opacification and the Lund-Mackay score ( ρ = 0.85, P < .001). Both percentage opacification and the Lund-Mackay score exhibited moderate correlations with the Lund-Kennedy score ( ρ = 0.58, P < .001, and ρ = 0.58, P < .001, respectively). The percentage osteitis correlated moderately with the Global Osteitis Scoring Scale ( ρ = 0.48, P < .001). CONCLUSIONS: Our quantitative processing of sinus CT images provides objective measures that correspond well to established visual scoring methods. While automation is a clear benefit here, validation may be needed in a prospective, multi-institutional setting. AI : artificial intelligence CRS : chronic rhinosinusitis CNN : convolutional neural network GOSS : Global Osteitis Scoring Scale LKS : Lund-Kennedy score LMS : Lund-Mackay score mHU : mean Hounsfield Units %OST : percentage of osteitis %SO : percentage sinus opacification SNOT-22 : SinoNasal Outcomes Test-22