An explainable machine learning approach for cervical cancer screening: decoding morphological diagnostic drivers

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Cervical cancer screening faces challenges due to pathologist shortages and diagnostic workload.
• Fatigue-induced misdiagnoses are a concern in manual screening processes.
• An interpretable machine learning (ML) tool was evaluated for its diagnostic capabilities.

Purpose of the Study:

• To assess the diagnostic capacity of an interpretable ML tool for cervical cancer screening.
• To evaluate the robustness of the ML tool across different patient subpopulations.
• To determine the operational efficiency gains from implementing the ML tool in a healthcare network.

Main Methods:

• A retrospective analysis of 5,000 women's liquid-based cytology (LBC) digital slides was conducted.
• The LBC-NET pipeline (ML tool) was compared against a panel of senior cytopathologists.
• Diagnostic safety was assessed across age, vaginal infections, and menopause; algorithmic transparency used SHAP and feature tracking; operational efficiency measured via cumulative distribution curves.

Main Results:

• The ML model achieved 92.5% sensitivity for high-grade squamous intraepithelial lesions or worse (HSIL+), outperforming manual methods (84.2%, p < 0.001) with an AUC of 0.94.
• No significant performance degradation was observed in patients with menopausal tissue atrophy or active vaginal infections.
• The ML system reduced slide-review time by 63.0%, increased daily screening capacity by 175%, and decreased specialist referrals by 45.2%.

Conclusions:

• The interpretable ML tool demonstrates high diagnostic accuracy and robustness for cervical cancer screening.
• The AI system significantly enhances operational efficiency, addressing pathologist workload challenges.
• This ML tool presents a reliable and fast clinical solution for mass cervical cancer screening.