Biological feature-based machine learning in histopathological images: a systematic review

Area of Science:

• Digital pathology and computational analysis of microscopic images.
• Artificial intelligence (AI) in medical diagnostics and pathology.
• Biomedical image analysis and machine learning applications.

Background:

• Deep learning (DL) models excel in histopathological slide analysis but lack interpretability.
• Handcrafted features from biological objects (nuclei, cells, tissues) offer better explainability.
• Machine learning (ML) with handcrafted features can complement DL for pathologist assistance.

Purpose of the Study:

• To systematically review the use of biological features in hematoxylin & eosin (H&E) microscopic images for medical questions.
• To identify feature categories, data sources, and medical applications in published literature.
• To assess methodological limitations and identify promising avenues for explainable AI in pathology.

Main Methods:

• Systematic literature review adhering to PRISMA guidelines.
• Analysis of 97 articles published between January 2005 and May 2025.
• Data extraction from PubMed, IEEE, and ACM databases focusing on feature types, sources, and medical questions.

Main Results:

• Identified three primary feature categories: texture/color, morphology, and topology.
• Features most commonly derived from segmented cells (80 studies) and tissues (28 studies).
• Features applied to seven medical questions: normal vs. diseased, subtyping, grading, phenotyping, object detection, prognosis, and treatment response.

Conclusions:

• Methodological limitations include interpretability challenges, data leakage, and small sample sizes.
• Domain-inspired feature engineering enhances explainability and specificity.
• Increased methodological rigor in feature engineering can improve AI model relevance and reliability in pathology.