Machine-learning-based evaluation of intratumoral heterogeneity and tumor-stroma interface for clinical guidance

Abstract

Assessment of intratumoral heterogeneity and tumor-host interaction within the tumor microenvironment is becoming increasingly important for innovative cancer therapy decisions because of the unique information it can generate about the state of the disease. However, its assessment and quantification are limited by ambiguous definitions of the tumor-host interface and by human cognitive capacity in current pathology practice. Advances in machine learning and artificial intelligence have opened the field of digital pathology to novel tissue image analytics and feature extraction for generation of high-capacity computational disease management models. A particular benefit is expected from machine-learning applications that can perform extraction and quantification of subvisual features of both intratumoral heterogeneity and tumor microenvironment aspects. These methods generate information about cancer cell subpopulation heterogeneity, potential tumor-host interactions, and tissue microarchitecture, derived from morphologically resolved content using both explicit and implicit features. Several studies have achieved promising diagnostic, prognostic, and predictive artificial intelligence models that often outperform current clinical and pathology criteria. However, further effort is needed for clinical adoption of such methods through development of standardizable high-capacity workflows and proper validation studies.