Supporting cardiac MRI generation with high fidelity digital twin

The generation of cardiac cine-MRI for both healthy and pathological cases is a highly relevant topic for simulation, training applications, and the creation of high-quality synthetic datasets. Despite recent advances in generative techniques, cardiac MRI remains particularly challenging due to the difficulty of producing realistic anatomical contexts and time sequences that adhere to the physical constraints imposed by blood flow dynamics. However, the emergence of high-fidelity computational models of the human heart paves the way for generating image sequences conditioned on the shapes and positions of cardiac chambers, as defined by the underlying physiological model. In this proof of concept, we present an artificial neural network derived from medical image deformation models. The network is trained on 360 cine-MRI volumes from the public M&M dataset and is designed to combine existing MRI sequences with conditional information generated by a high-fidelity cardiac digital twin, enabling the generation of both short-axis 3D MRI sequences.