Deep Learning-Based 3D Leukocyte Differentiation Using Label-Free Higher Harmonic Generation Microscopy

Abstract

Both in clinical practice and translational research, cell differentiation of leukocytes provides important information used for diagnostics or insights into pathophysiological mechanisms. The current gold-standard method for bronchoalveolar lavage fluid (BALF) analysis involves histochemical staining of cytospins, followed by manual morphological quantification. However, this approach is labor-intensive, time-consuming, and highly operator-dependent, limiting its efficiency and throughput. This study proposes a deep learning framework for rapid, automated 3D leukocyte differentiation using label-free higher harmonic generation microscopy (HHGM). 3D leukocyte characterization was performed with label-free HHGM. Two deep learning models, ResNet 3D-50 and Vision Transformer (ViT) 3D, were trained, validated and tested for leucocyte differentiation on both BALF and blood fraction samples from 14 interstitial lung disease (ILDs) and 12 acute respiratory distress syndrome (ARDS) patients. Deep-learning model-prediction and cytospin analysis were performed by separate investigators. Results were compared using Bland-Altman analysis. The deep learning algorithm achieved >96% accuracy in quantifying neutrophils, eosinophils, lymphocytes, and macrophages/monocytes. Bland-Altman analysis showed mean differences of less than 3% between cytospin analysis and the deep learning based approach across all cell types. By integrating the label-free imaging capabilities of HHGM with deep learning, this study established a fast, accurate and high-throughput leukocyte differentiation in fresh BALF and blood samples. By significantly improving efficiency and reproducibility, this technology has the potential to transform clinical workflows and advance precision medicine.