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

Area of Science:

• Digital pathology
• Artificial intelligence in diagnostics
• Cervical cancer screening

Background:

• Histopathologic grading of cervical biopsies has interobserver variability, especially at the CIN2+ threshold.
• Deep learning systems offer potential to improve diagnostic accuracy in histopathology.
• Standardizing cervical cancer screening requires reliable diagnostic tools.

Purpose of the Study:

• To evaluate a deep learning system (EagleEye) for detecting CIN2+ on H&E whole-slide images.
• To compare EagleEye's performance with independent pathologists and an AI-assisted workflow.
• To preliminarily assess spatial correspondence between AI heatmaps and p16 staining.

Main Methods:

• Digitized 99 archived cervical biopsies (Normal, CIN1, CIN2, CIN3, ACIS) in a spectrum-balanced design.
• Compared diagnoses from original sign-out (P1), a second pathologist (P2), EagleEye alone (EE), and AI-assisted read (EE + P2).
• Evaluated agreement using Cohen's κ and sensitivity/specificity; assessed p16-to-AI spatial correspondence in 30 cases.

Main Results:

• EagleEye achieved 93.3% sensitivity and 71.8% specificity for CIN2+ detection versus P1.
• The AI-assisted workflow (EE + P2) improved P1's sensitivity to 83.8% while maintaining 100% specificity.
• EagleEye flagged potential squamous cell carcinoma (SCC) in CIN3 cases, confirmed by expert review; p16-to-AI correspondence was ≥70% in 73.3% of cases.

Conclusions:

• EagleEye demonstrates high sensitivity for CIN2+ detection and substantial agreement with expert readers in an AI-assisted workflow.
• The primary benefit of AI assistance was improved identification of treatment-relevant lesions, particularly near the CIN1/CIN2 boundary.
• Final diagnostic decisions remain with the pathologist, leveraging AI as a supportive tool.