Digital Pathology with AI for Cervical Biopsies: Diagnostic Accuracy at the CIN2+ Threshold

Background/Objectives: Histopathologic grading of cervical biopsies is subject to interobserver variability, particularly at the CIN2+ treatment threshold. We evaluated a deep learning system (EagleEye) for detecting CIN2+ (CIN2, CIN3, ACIS, invasive carcinoma) on hematoxylin-eosin (H&E) whole-slide images (WSIs) and compared its performance with independent pathologists, including an AI-assisted workflow. Spatial correspondence with p16 staining was preliminarily assessed. Methods: Ninety-nine archived cervical punch biopsies from a single university hospital, originally diagnosed as Normal (n = 19), CIN1 (n = 20), CIN2 (n = 20), CIN3 (n = 20), or adenocarcinoma in situ (ACIS; n = 20), were digitized in a deliberately spectrum-balanced design. The original sign-out (P1), a second gynecologic pathologist (P2, microscope and digital), EagleEye alone (EE), and an AI-assisted read (EE + P2) served as diagnostic conditions. Outcomes were dichotomized as Results: Agreement between P1 and EagleEye was moderate (κ = 0.67), while P2 showed high internal consistency (κ = 0.86) and good agreement with P1 (κ = 0.78). Using P1 as reference, EagleEye detected CIN2+ with 93.3% sensitivity and 71.8% specificity. When the AI-assisted consensus (EE + P2) was used as an augmented internal comparator, P1 showed 83.8% sensitivity and 100% specificity, indicating that the human-in-the-loop workflow identified additional CIN2+ cases that had been signed out as Conclusions: In this single-centre, spectrum-balanced cohort, EagleEye achieved high CIN2+ sensitivity and substantial agreement with expert readers in a human-in-the-loop workflow. The main added value was internal case-finding of treatment-relevant lesions when AI assistance was available, while final diagnoses remained with the pathologist.