The tradeoff of cell classifier error rates

Area of Science:

• Gynecologic cytopathology
• Biomedical image analysis
• Statistical modeling

Background:

• Specimen false negative errors in gynecologic cytodiagnosis are considered more severe than false positives.
• Optimizing parameters for individual cell classification to balance error rates is not straightforward.
• Cervical cytology involves analyzing samples with a low proportion of abnormal cells.

Purpose of the Study:

• To analyze the trade-offs between cell false positive and false negative rates in gynecologic cytodiagnosis.
• To determine optimal parameter settings for cell classifiers in the context of specimen-level classification.
• To calculate the required sample size for achieving specific overall performance metrics.

Main Methods:

• Developing a model for a specimen classifier that follows a cell classifier.
• Analyzing the impact of cell error rates on overall specimen diagnosis accuracy.
• Simulating performance based on typical abnormal cell proportions in cervical cytology.

Main Results:

• The analysis demonstrates that for cervical cytology, cell false positive rates must be significantly minimized.
• Achieving low cell false positive rates necessitates accepting a substantial rate of missed abnormal cells.
• The study quantifies the sample size needed for desired overall diagnostic performance.

Conclusions:

• In cervical cytology, prioritizing the reduction of cell false positives is essential, despite the inherent risk of missing some abnormal cells.
• The findings provide guidance on adjusting classifier parameters to manage error rates effectively in gynecologic cytodiagnosis.
• This approach ensures that the more critical specimen-level false negative errors are controlled.