Cytoplasm and nuclei as a basis for Bethesda cell cluster classification in cervical smears

Abstract

Population screening in the form of cervical smears was introduced to reduce cervical cancer morbidity. However, the manual evaluation of cervical smears is a labour-intensive and meticulous task. This challenge has led to the development of various computer-aided cell identification methods aimed at improving the quality of evaluations and reducing the time required for slide analysis. These supportive tools for pathologists should adhere to the Bethesda classification system for cell types, facilitating integration with established clinical practices. Despite advances, the automatic classification of clustered cells in cervical smears remains a significant challenge for both automated and semiautomated methods that utilize classical image processing and deep learning techniques.

This study introduces a novel method for fragmenting clusters to improve the classification of clustered cells in digital images of Papanicolaou smears. The proposed method integrates explainable AI and marker-guided watershed segmentation ensuring both interpretability and versatility of an overall pipeline for diagnostician support in smear evaluation. Validation of the method was performed on a recently developed Papanicolaou cytology Bialystok dataset, as well as the publicly available CRIC dataset, demonstrating the method’s generalizability across different datasets.

The cell classification pipeline, enhanced by the proposed declustering method, improved the overall harmonic mean of recall and precision (F1 score) by 13.27 percentage points compared with the results obtained without this additional processing. The improvement in classifying the most critical cell type according to the Bethesda system (HSIL cell class) was even more significant, with an increase of 35.72 percentage points compared with classifying the entire cluster.