Staining, magnification, and algorithmic conditions for highly accurate cell detection and cell classification by...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Cytodiagnosis

Background:

• Deep learning models are increasingly explored for automated cell detection and classification in cytodiagnosis.
• Challenges remain in developing general-purpose cytology models due to variations in magnification, staining, and false positives.

Purpose of the Study:

• To evaluate the performance of deep learning models for general-purpose cytology applications.
• To investigate the impact of magnification and staining methods on model accuracy.
• To address challenges like false positives in AI-based cytodiagnosis.

Main Methods:

• Developed deep learning models using the You Only Look Once, version 8 (YOLOv8) algorithm.
• Utilized 11 types of human cancer cell lines with Papanicolaou and May-Grünwald-Giemsa (MGG) staining.
• Assessed detection and classification rates across different cell types, staining methods, and magnifications.

Main Results:

• All models achieved classification rates exceeding 95.9%.
• Highest detection rates were 92.3% for Papanicolaou-stained and 91.3% for MGG-stained models.
• Object detection and instance segmentation models reached 94.6% and 91.7% detection rates, respectively, for 11 cell types with Papanicolaou staining.

Conclusions:

• YOLOv8 artificial intelligence technology demonstrates sufficient performance for clinical cytodiagnosis screening and cell classification.
• Further research on clinical specimens is essential to validate YOLOv8 efficacy and address specific cytology challenges.