Evaluating Cellularity Estimation Methods: Comparing AI Counting with Pathologists' Visual Estimates

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in oncology

Background:

• Next-generation sequencing (NGS) drives precision medicine by identifying cancer driver gene alterations.
• Accurate genetic testing relies on sufficient tumor cells, necessitating precise tumor content ratio (TCR) evaluation.
• Pathologist variability in TCR estimation from H&E images presents a significant challenge.

Purpose of the Study:

• To develop and validate an AI model for accurate TCR calculation from H&E-stained lung cancer images.
• To compare the accuracy of AI-based TCR estimation against pathologist visual assessment.
• To assess the robustness and inter-institutional consistency of the developed AI model.

Main Methods:

• Established a "gold standard" TCR by exhaustive cell labeling by three pathologists on 41 lung cancer cases.
• Developed a compact, fast, fully convolutional neural network AI model for cell detection and classification.
• Compared TCR accuracy between 13 pathologists and the AI model against the gold standard.

Main Results:

• The AI model achieved 92% cell detection and 84% classification accuracy.
• AI-based TCR calculation showed significantly lower error compared to pathologist visual assessment (p<0.05).
• The AI model demonstrated greater inter-institutional robustness than the average pathologist assessment.

Conclusions:

• AI models significantly improve the accuracy of tumor cellularity assessments in clinical settings.
• Enhanced TCR accuracy via AI facilitates more efficient and reliable genetic testing.
• The adoption of robust AI tools promises improved patient outcomes through precise cancer treatment.