Machine Learning-Based Label-Free Macrophage Phenotyping in Immune-Material Interactions

Abstract

The rapid advancement of implantable biomedical materials necessitates a comprehensive understanding of macrophage interactions to optimize implant immunocompatibility. Macrophages, key immune regulators, exhibit phenotypic plasticity by polarizing into pro-inflammatory (M1) or anti-inflammatory (M2) subtypes. Conventional phenotyping techniques, such as flow cytometry and immunostaining, provide insights but have limitations related to fixation and endpoint analysis. This study presents a high-throughput, label-free macrophage phenotyping approach integrating AI-driven image classification with quantitative phase imaging (QPI). THP-1-derived macrophages were differentiated into M0, M1, M2a, and M2c phenotypes, and their morphological and refractive index properties were analyzed using QPI. Although QPI alone could not fully distinguish phenotypes, deep learning models, including GoogLeNet, ShuffleNet, VGG-16, and ResNet-18, were evaluated, with ResNet-18 achieving over 90% accuracy. Additionally, macrophage responses to collagen coatings (types I, III, and IV) were assessed using machine learning-based phenotyping and cytokine profiling. Collagen I induced an M1 response, collagen III supported a balanced M1/M2 profile, and collagen IV promoted a controlled immune environment. These findings demonstrate the potential of AI-driven QPI as a non-invasive tool for macrophage characterization, offering insights into biomaterial immunocompatibility and informing implant design strategies.