Toward Interpretable Cell Image Representation and Abnormality Scoring for Cervical Cancer Screening Using Pap Smears

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Healthcare

Background:

• Cervical cancer screening is vital for early detection but is labor-intensive.
• Deep learning models show promise for automating Pap smear analysis.
• Class imbalance and high labeling costs in healthcare data necessitate alternative training strategies.

Purpose of the Study:

• To develop an explainable deep learning method for cervical cell representation using one-class classification.
• To enable abnormality detection in Pap smear images without using abnormal samples during training.
• To localize and interpret detected cell abnormalities effectively.

Main Methods:

• Utilized variational autoencoders for one-class classification of cervical cytology images.
• Developed a scoring system for cell abnormality based on learned representations.
• Employed agglomerative clustering with a novel cross-entropy difference metric for abnormality localization.

Main Results:

• Achieved an Area Under the Operating Characteristic Curve (AUC) of 0.908±0.003 for discriminating squamous cell carcinoma (SCC) from normal cells.
• Achieved an AUC of 0.920±0.002 for discriminating high-grade squamous intraepithelial lesion (HSIL) from normal cells.
• Demonstrated improved V-measure and homogeneity scores compared to other clustering methods, enhancing abnormality region isolation.

Conclusions:

• The proposed one-class classification method effectively detects cervical cell abnormalities without requiring abnormal training data.
• The explainable deep representations and novel localization metric aid in interpreting screening results.
• The model shows robust performance on external datasets, indicating its potential for real-world application in cervical cancer screening.