Scrutinizing high‐risk patients from ASC‐US cytology via a deep learning model
BACKGROUND
Atypical squamous cells of undetermined significance (ASC‐US) is the most frequent but ambiguous abnormal Papanicolaou (Pap) interpretation and is generally triaged by high‐risk human papillomavirus (hrHPV) testing before colposcopy. This study aimed to evaluate the performance of an artificial intelligence (AI)‐based triage system to predict ASC‐US cytology for cervical intraepithelial neoplasia 2+ lesions (CIN2+).
METHODS
More than 60,000 images were used to train this proposed deep learning‐based ASC‐US triage system, where both cell‐level and slide‐level information were extracted. In total, 1967 consecutive ASC‐US Paps from 2017 to 2019 were included in this study. Histological follow‐ups were retrieved to compare the triage performance between the AI system and hrHPV in 622 patients with simultaneous hrHPV testing.
RESULTS
In the triage of women with ASC‐US cytology for CIN2+, our system attained equivalent sensitivity (92.9%; 95% confidence interval [CI], 75.0%‐98.8%) and higher specificity (49.7%; 95% CI, 45.6%‐53.8%) than hrHPV testing (sensitivity: 89.3%; 95% CI, 70.6%‐97.2%; specificity: 34.3%; 95% CI, 30.6%‐38.3%) without requiring additional patient examination or testing. Additionally, the independence of this system from hrHPV testing (κ = 0.138) indicated that these 2 different methods could be used to triage ASC‐US as an alternative way.
CONCLUSION
This de novo deep learning‐based system can triage ASC‐US cytology for CIN2+ with a performance superior to hrHPV testing and without incurring additional expenses.;