Robust Cell-Level Classification for Liquid-Based Cervical Cytology Using Deep Transfer Learning: A Multi-Source...

Area of Science:

• Digital pathology
• Machine learning in healthcare
• Cytopathology

Background:

• Automated cervical cytology analysis using deep learning faces challenges with model generalization due to domain shifts from different scanners and labs.
• Variations in color, texture, and morphology across datasets hinder the performance of AI models in Pap smear analysis.

Purpose of the Study:

• To develop a robust cell-level classification framework for liquid-based Pap smear cytology using deep transfer learning.
• To enhance model generalization across heterogeneous acquisition conditions by integrating multi-source datasets.

Main Methods:

• A multi-source dataset was created by combining public repositories (SIPaKMeD, Herlev, CRIC Cervix) with a proprietary cohort of Whole Slide Images (WSIs).
• Cell-centered image patches were standardized for training and evaluating CNN models (ResNet50, EfficientNetB0, VGG16).
• Model robustness was assessed by systematically analyzing performance across different data-source combinations and acquisition variations.

Main Results:

• ResNet50 achieved the best performance on the independent test set with 0.91 accuracy and 0.91 macro-F1 score.
• Integrating proprietary multi-center data significantly improved robustness to scanner-induced variations compared to using public data alone.
• The study demonstrated that diverse data integration mitigates domain shift in cell-level cervical cytology classification.

Conclusions:

• A robust cell-level classification framework for liquid-based Pap smear cytology was successfully developed using deep transfer learning.
• Combining diverse datasets, including proprietary multi-center data, enhances model robustness against acquisition variability.
• The proposed classifier serves as a foundational component for future end-to-end Whole Slide Image screening pipelines in cervical cancer screening.