Relationship Between Artificial Intelligence-Based Cell Detection and Cytomorphological Variations Induced by Cell-Processing Solutions: Usefulness of Data Augmentation in Artificial Intelligence Cytology.

Abstract

Introduction: Variations in cytomorphology due to differences in specimen preparation conditions hinder the implementation of artificial intelligence (AI) in cytology. In addition, small-scale research and insufficient datasets pose challenges. In this study, we aimed to examine the relationship between cytomorphological variations induced by cell-processing solutions and AI-based cell detection accuracy, and to demonstrate the usefulness of data augmentation in AI cytology.

Methods: Samples of untreated MKN45 human gastric cancer cells and cells treated with four different cell-processing solutions were used to prepare the specimens. These specimens were subjected to Papanicolaou staining, and the areas and hue, saturation, and brightness (HSB) values of the nucleus and cytoplasm were analyzed mathematically and statistically. Deep learning (DL) models were developed with and without data augmentation, and AI-based cell detection rates were evaluated.

Results: Heparin sodium solution-treated (Hep) cells showed obvious differences from other groups and presented significant differences compared to control (Cont) cells in the analysis of areas and HSB values of the nucleus and cytoplasm. The AI-based cell detection rate of Hep cells was also significantly lower than that of Cont cells. The use of the DL model with data augmentation improved the AI-based cell detection rate for all samples.

Conclusion: We identified the key cytomorphological features that AI focuses on when recognizing cells and demonstrated that data augmentation is an effective technique for improving AI-based cell detection accuracy.