Learning Cellular Phenotypes through Supervision

Area of Science:

• Computational pathology
• Biomedical image analysis
• Machine learning in medicine

Background:

• Image-based cell phenotyping is crucial for computational pathology but faces challenges in robustness and biological relevance.
• Ensuring cell cluster properties are insensitive to experimental variations and that phenotypes support clinical reporting are key issues.

Purpose of the Study:

• To compare the robustness and clinical relevance of self-supervised versus supervised features for cell phenotyping.
• To demonstrate the application of model explainability (Shapley values) for identifying disease-relevant cellular phenotypes.

Main Methods:

• Phenotype consistency was evaluated using both self-supervised and supervised features.
• Case classification was used to analyze feature set relevance for clinical diagnosis.
• Shapley values were employed for model explainability and identifying important cellular phenotypes.

Main Results:

• Supervised features demonstrated improved accuracy (7.4%) in case classification for myeloproliferative neoplasms compared to self-supervised features.
• Shapley values successfully identified disease-relevant cellular phenotypes and their importance.

Conclusions:

• Supervised features offer enhanced accuracy for cell phenotyping in computational pathology, particularly for diagnosing hematologic disorders.
• Integrating model explainability aids in discovering clinically significant cellular phenotypes and improving diagnostic models.