Biomechanically Regularized Deep Deformable Registration for CT and US Fusion
Snigdha Mohanty, Sara Yavari, Sarada Prasad Dakua · Electronics Letters · 2025
ABSTRACT Accurate registration of liver computed tomography (CT) and ultra‐sound (US) images is crucial for image‐guided interventions, enabling precise navigation and treatment planning. However, aligning these multimodal images is challenging due to differences in imaging characteristics, patient positioning, and organ deformation. This paper proposes a novel biomechanically regularized deep deformable registration framework for liver CT and US fusion. Our approach leverages deep learning to model complex deformation fields while incorporating biomechanical priors derived from liver anatomy and tissue properties to enhance the physical plausibility and robustness of the registration process. We utilize a convolutional neural network to predict the deformation field between the CT and US images, guided by a biomechanical regularization term that penalizes unrealistic deformations. The framework is trained on a dataset of paired liver CT and US images, incorporating both synthetic and real‐world deformation scenarios. The results show that the values of mean square error, normalized cross‐correlation, structural similarity, mutual information, feature similarity index, and mean absolute error are 1.2132, 0.9932, 0.9907, 0.913, 0.9971 and 1.39 1.02, respectively. This work represents a step forward in bridging the gap between multimodal image registration and biomechanically informed modelling, advancing precision in image‐guided liver interventions.