Deep learning-enabled tissue engineering scaffold classification using cell morphology

The physical properties and architecture of biomaterial scaffolds regulate cell morphology and function; however, evaluating these effects is often slow, low-throughput, and destructive, thereby limiting scalable biomaterial design and optimization. We investigated whether quantitative analysis of single-cell three-dimensional (3D) morphology could provide a non-destructive surrogate readout of scaffold type. We analyzed 969 high-resolution 3D reconstructions of human bone marrow stromal cells (hBMSCs) from a publicly available National Institute of Standards and Technology (NIST) dataset. Cells were cultured under ten conditions, grouped by scaffold type: flat two-dimensional (2D) surfaces, fibrous 3D scaffolds, porous 3D sponge scaffolds, and 3D hydrogel scaffolds. Stiffness was treated as contextual metadata rather than a direct classification label. We developed two complementary classifiers. The first was a radiomics pipeline that extracted hand-crafted shape descriptors and used a feed-forward neural network, achieving a best test accuracy of 73.2%. The second employed a ray-tracing pipeline that transformed 3D cell structure into 2D distance-map projections for convolutional neural network (CNN) analysis, achieving 72.7% accuracy. The radiomics model offered greater interpretability, whereas the ray-tracing model captured more subtle morphological features, demonstrating complementary strengths. These findings suggest that single-cell 3D morphology can encode scaffold type under the controlled conditions examined here and may offer a non-destructive basis for biomaterial identification that warrants further validation. Such a framework could potentially accelerate scaffold design, optimization, and automated quality control in tissue engineering applications. Statement of Significance: Development of advanced biomaterial scaffolds is constrained by characterization methods that are often destructive, labor-intensive, and poorly suited for high-throughput optimization. Here, we present a morphology-based computational framework in which single-cell 3D actin and nuclear morphology is used as a quantitative readout of scaffold-associated cellular phenotype. The method integrates two complementary approaches: an interpretable radiomics pipeline based on 3D geometric descriptors and a ray-tracing pipeline that converts 3D cell morphology into 2D distance maps for convolutional neural network classification. Using nearly 1000 reconstructed human bone marrow stromal cells cultured on flat, fibrous, porous, and hydrogel scaffold conditions, both pipelines classified scaffold architecture with test accuracies of approximately 73%. These findings demonstrate that 3D cell morphology encodes scaffold-associated information and support computational morphotyping as a scalable strategy for biomaterial screening, scaffold quality assessment, and phenotype-guided scaffold prioritization in regenerative medicine.