Introduction: Because MR does not involve ionizing radiation and may be used without contrast, recently there has been a trend towards monitoring aneurysms primarily with MRA and then use CTA as needed. Our goal is to establish a machine learning technique for aneurysm evaluation that can generate images with comparable image quality to CTA based on time of flight (TOF) MRA.
Methods: TOF MRA and CTA images acquired less than 2 weeks apart were used. A set of clinical TOF MRA and CTA were used for training and five additional sets were used for testing. TOF was obtained with gradient recalled echo (TR = 33ms, TE=3.2ms) on 1.5T scanners. CTA scans were acquired using 64 slice scanner (pitch=1.0, reconstruction kernel H30f). During the training stage, cross modality registration was performed to maximize mutual information between TOF and the corresponding CTA. To generate machine learning angiography (MLA) images, the input TOF were divided into sliding patches and their coefficient was estimated and used as an input to the neural network to infer voxelwise CTA values. Figure: images used for learning (aneurysm, right ICA; greyscale, CTA; red, TOF MRA).
Results: The MLA images enhanced visibility of artery structures from the original TOF. High geometric agreement was achieved between MLA images and CTA. In comparison to CTA, MLA images had sub-mm Hausdorff distance and pairwise distance of 0.24±0.12mm for all aneurysm localities for all tested image sets. With an aneurysm growth threshold defined as size change more than 0.6 mm in CTA, MLA provides sufficient resolution for cross modality follow up comparisons.
Conclusions: This pilot study has demonstrated promise to enhance artery visibility and interpretation from MRA using machine learning techniques to yield CTA-like images. The performance of the current method is affected by the dependence of flow enhancement with respect to imaging planes in TOF and requires consistent imaging protocols for training and testing subjects.