Deep Learning Based Registration of Dynamic Myocardial Perfusion CT Using VoxelMorph

Abstract

This work presents a deep learning-based approach for the registration of dynamic 2D Computed Tomography (CT) images of myocardial perfusion using VoxelMorph framework. The method aligns cardiac CT slices by training a convolutional neural network (CNN) to learn spatial transformations between pairs of moving and fixed images. The model was trained and validated on a dataset comprising 118 patients diagnosed with or suspected of having coronary artery disease and/or aortic valve insufficiency. A supervised learning strategy was used, combining mean squared error (MSE) with a smoothness regularization term to guide the prediction of deformations. Compared to the baseline model without regularization, the proposed method improved the Dice Similarity Coefficient (DSC) from 0.86 to 0.90 and increased the Peak Signal to Noise Ratio (PSNR) from 27.2 to 29.8 dB, reflecting a more accurate spatial alignment. The efficiency of the experimental results indicates the potential of the model for real time cardiac image alignment.