Ma, S; Lin, Z; Du, X; Hu, Y; Min, Z; (2026) The Impact of Biomechanical Quantities on PINNs-Based Medical Image Registration. In: Ni, D and Noble, A and Huang, R and Xue, W, (eds.) Simplifying Medical Ultrasound. ASMUS 2025. (pp. pp. 98-108). Springer: Cham, Switzerland.

Text PINNS_Registration_submit_asmus.pdf - Accepted Version Access restricted to UCL open access staff until 28 September 2026. Download (992kB)

Abstract

Biomechanical constraints are increasingly being adopted in medical image registration to enforce physically plausible soft-tissue deformations. When formulating loss functions for physics-informed neural networks (PINNs)-based registration, a central design choice is whether the network should explicitly predict biophysical model parameters or implicitly infer them by enforcing their governing equations. In this study, we introduce four physics-informed regularization strategies for biomechanics-informed non-rigid point set registration: (i) stress prediction, (ii) strain prediction, (iii) stress–strain prediction, and (iv) deformation prediction. We provide a promising direction to understand and potentially utilised prior biomechanical knowledge to modern data-driven approach for motion modelling and medical image registration. On the simulation dataset, the deformation prediction strategy achieved the strongest performance, reducing the root mean square error (RMSE) from 1.748 ± 0.912 mm (FPT, without PINN) to 0.219 ± 0.057 mm (-87.5%) and chamfer distance from 1.260 ± 0.913 mm to 0.194 ± 0.057 mm, without generating negative jacobian determinants. Other PINNs strategies also yielded competitive results, stress (RMSE 0.264 ± 0.098 mm), strain (0.305 ± 0.116 mm), and stress–strain (0.242 ± 0.147 mm), all of which consistently outperformed conventional baselines such as BCPD (RMSE 2.326 ± 5.811 mm) and GMM-FEM (1.528 ± 0.912 mm). On the clinical dataset, the deformation prediction strategy again obtained the lowest target registration error (TRE 4.924 ± 1.542 mm, -16.6% vs. FPT), while maintaining a chamfer distance of 2.125 ± 0.291 mm and only negligible foldings (0.278 ± 0.312% negative Jacobians). The other PINNs variants, stress (TRE 5.694±1.780 mm), strain (5.359±1.746 mm), and stress–strain (5.912±2.008 mm), showed similar improvements over non-PINNs approaches, but remained slightly inferior to deformation prediction. The deformation prediction strategy was statistically significant against all methods (paired p<0.05), except BCPD on the simulation dataset. The source codes have been released at https://github.com/Msx00/PINNs-for-Point-Set-Registration.git.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as CSVJSONXML

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item