Abstract
Mesenchymal stem cells (MSCs) are widely applied in regenerative medicine, but conventional osteogenic induction assays are time-consuming and rely on destructive endpoint measurements. Here, we propose OsteoNet, a deep learning framework that predicts osteogenic differentiation from bright-field images and generates an Osteogenic Score (OsScore) reflecting differentiation dynamics. The predictive performance of OsteoNet was evaluated across multiple time points using an independent test set, achieving an AUC of 0.94 on day 0 and 0.98 on day 5, demonstrating robust early-stage detection capability. The OsScore increased progressively with induction time and showed strong positive correlations with both early and late osteogenic markers, including RUNX2, OCN, and OSX, at the RNA and protein levels. Morphological analysis of immunofluorescence images further confirmed significant increases in cell size during early differentiation, supporting the model's sensitivity to subtle morphological cues. Collectively, OsteoNet enables non-invasive, quantitative, and early monitoring of osteogenic differentiation in hAMSCs, offering a powerful tool to accelerate research and reduce reliance on destructive endpoint assays.