Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
BP: Fachverband Biologische Physik
BP 7: Poster Session I
BP 7.24: Poster
Montag, 9. März 2026, 15:00–17:00, P5
Deep learning architecture combining Vision Transformers and U-Nets for robust traction force microscopy — •Yunfei Huang1, Elena Van der Vorst2,3, and Benedikt Sabass1,2,3 — 1Faculty of Physics, Technical University Dortmund, Dortmund, 44227, Germany — 2Faculty of Physics and Center for NanoScience, Ludwig Maximilian University of Munich, Munich, 80752, Germany — 3Department of Veterinary Science, Ludwig Maximilian University of Munich, Munich, 80752, Germany
Traction force microscopy (TFM) quantifies cellular forces on the extracellular matrix. Although deep learning has advanced TFM analysis, challenges remain in achieving reliable inference across spatial scales, estimating uncertainty, and integrating biological information such as cell type. In this study, we propose a robust deep learning architecture, ViT+UNet, which integrates a U-Net with a Vision Transformer. Our results show that this hybrid model outperforms either U-Net or ViT alone in predicting traction force fields. In addition, ViT+UNet achieves superior generalization across a wide range of scales, which allows one to use the algorithm for TFM with different setups and equipment. We extend the model with an uncertainty estimation module that enables simultaneous prediction of traction forces and confidence levels. Incorporating cell-type information further improves accuracy. Simulated results show that the algorithm effectively reconstructs 3D traction fields in non-linear elastic matrices.
Keywords: Traction force microscopy; Deep learning; Vision Transformers and U-Nets; Cell-type; Uncertainty