Journal Article Review Epublish
PubMed: 41623734 DOI: 10.1016/j.jpi.2025.100539
Biological feature-based machine learning in histopathological images: a systematic review.
J Pathol Inform 20:100539
Authors: Stéphane Treillard, Robin Schwob, Sandrine Mouysset, Pierre Brousset, Sylvain Cussat-Blanc, Camille Franchet
Affiliations
- CHU de Toulouse, Toulouse, France. Université Toulouse Capitole, Toulouse, France. Institut de Recherche en Informatique de Toulouse - CNRS UMR5505, France.
- CHU de Toulouse, Toulouse, France. Institut de Recherche en Informatique de Toulouse - CNRS UMR5505, France. Université de Toulouse Jean Jaurès, Toulouse, France.
- Institut de Recherche en Informatique de Toulouse - CNRS UMR5505, France. Université de Toulouse, Toulouse, France.
- Université de Toulouse, CHU Toulouse, IUCT-Oncopole, Toulouse, France.
- Université Toulouse Capitole, Toulouse, France.
- Institut de Recherche en Informatique de Toulouse - CNRS UMR5505, France.
- Institut Universitaire de France, Paris, France.
- Université de Toulouse, Oncopole Claudius Regaud, IUCT-Oncopole, Toulouse, France.
Abstract
Digital pathology has recently led to significant advancements in the field of microscopic image analysis, particularly regarding the increasing use of Deep Learning methods. These models represent the state-of-the-art in histopathological slide analysis, but Deep Learning features remain difficult to interpret, despite recent developments in post hoc explainability frameworks. In contrast, features extracted from biological objects-such as nuclei, cells or tissues-are supposed to be more grounded in pathologists' a priori knowledge. Accordingly, Machine Learning based on handcrafted features represents another paradigm of explainability and may stand as a complementary method to Deep Learning to assist pathologists. In order to perceive how biological features have been used in hematoxylin & eosin microscopic images to address medical questions, we conducted a systematic review of articles published from January 2005 to May 2025, adhering to PRISMA guidelines. A total of 97 articles were analyzed from the PubMed, IEEE, and ACM databases. Three primary categories of features-texture/color, morphology and topology-were both identified and thoroughly described. These features were most frequently derived from segmented cells and tissues in 80 and 28 studies, respectively. They were used to address seven types of medical questions: "normal vs diseased", disease subtyping, tumor grading, phenotyping, object detection, prognosis and treatment-response prediction. We discussed methodological and reporting limitations of these studies, highlighting the difficulty to assess the potential impact of such methods. Among the most common concerns, we found features difficult to interpret, data leakage, and inadequate sample sizes. Nevertheless, we also focused on promising domain-inspired feature engineering that provides better explainability and specificity. This kind of features associated with more methodological rigor may increase the relevance and reliability of AI models, and also raise new research avenues in pathology.
MeSH
Cells, Pathologists, Pathology, Machine Learning, Learning, Tissues