Model-based cap thickness and peak cap stress prediction for carotid MRI

Subvoxel vessel wall thickness measurements of the intracranial arteries using a convolutional neural network

Vessel wall thickening of the intracranial arteries has been associated with cerebrovascular disease and atherosclerotic plaque development. Visualization of the vessel wall has been enabled by recent advancements in vessel wall MRI. However, quantifying early wall thickening from these MR images is difficult and prone to severe overestimation, because the voxel size of clinically used acquisitions exceeds the wall thickness of the intracranial arteries. In this study, we aimed for accurate and precise subvoxel vessel wall thickness measurements. A convolutional neural network was trained on MR images of 34 ex vivo circle of Willis specimens, acquired with a clinically used protocol (isotropic acquired voxel size: 0.8 mm). Ground truth measurements were performed on images acquired with an ultra-high-resolution protocol (isotropic acquired voxel size: 0.11 mm) and were used for evaluation. Additionally, we determined the robustness of our method by applying Monte Carlo dropout and test time augmentation. Lastly, we applied our method on in vivo images of three intracranial aneurysms to measure their wall thickness. Our method shows resolvability of different vessel wall thicknesses, well below the acquired voxel size. The method described may facilitate quantitative measurements on MRI data for a wider range of clinical applications.