Mechanical interpretation of flow phenomena in vivo is crucial for understanding pathological changes such as vascular disorders. Recent advances in fluid data assimilation combining clinical medical imaging with computational fluid dynamics (CFD) have enabled the reconstruction of flow fields faithful to observed data. While variational data assimilation effectively estimates physical parameters using CFD, its high computational cost limits clinical application. Physics-informed neural networks (PINNs), which embed physical laws into deep learning, offer a promising alternative, especially when combined with transfer learning (fine-tuning). However, applying fine-tuning directly to patient-specific geometries remains challenging due to the variability of vascular shapes. To address this difficulty, we propose a fine-tuning approach incorporating coordinate transformations. The method’s effectiveness is demonstrated on two-dimensional steady flow problems with simple geometries.
Ueda et al. (Wed,) studied this question.