Background: Patients who present with suspected stroke most commonly receive a CT scan for initial evaluation, due to lower cost and shorter acquisition time compared to MRI. However, MRI has been shown to be more accurate than CT for diagnosing stroke and can provide additional information that can be useful for have other advantages over CT scans for the purpose of diagnosis, treatment, and prognosis of stroke.
Aims: The purpose of this study was to investigate the use of deep learning to generate a synthetic MRI from a patient’s CT scan which could be used in a clinical setting of suspected stroke.
Methods: Eight deep learning models were implemented including 7 UNet based models, and a CycleGAN model. These models were trained on a dataset of 181 stroke patients who underwent both CT and MRI scans. The performance of these models was evaluated both visually and using quantitative metrics to assess their ability to generate synthetic MRIs. The synthetic MRIs were further evaluated on three clinical tasks: lesion segmentation, brain tissue segmentation, and registration to an MRI.
Results: All eight models were capable of generating synthetic MRIs that showed similarities to the true MRIs. The base 3D UNet model outperformed the other models in terms of quantitative metrics and most clinical tasks, while the CycleGAN model had the weakest performance. Several models were able to accurately translate CT lesions to synthetic MRIs, a pre-trained segmentation model could automatically segment these lesions.
Conclusion: The use of deep learning to generate synthetic MRIs for stroke patients appears to offer a viable means of obtaining the benefits of an MRI scan for those who only receive a CT scan. However, further research is required to evaluate the feasibility of synthetic MRIs in clinical practice.