Deep Learning-Based generation of synthetic CT from MR images for craniosynostosis planning

Craniosynostosis is a rare congenital defect caused by the premature fusion of one or more cranial sutures. This untimely cranial ossification hinders correct brain development. Its clinical diagnosis and treatment planning usually rely on Computed Tomography (CT), a potentially harmful imaging technique for young infants. It is with the intent of avoiding the use of ionizing radiation in this clinical pipeline that this work studies how feasible it is to resort to alternative non-detrimental imaging techniques such as Magnetic Resonance Imaging (MRI). We evaluate the performance of neural network generators trained on Generative Adversarial Networks in the MRI-to-CT translation task. We train nine generative models on 25 paired MR-CT medical scans, and validate and test their performance on 8 and 4 paired images, respectively. The results are promising both from qualitative and quantitative standpoints, particularly those of the models trained directly on 3D data. Results demonstrate that it is feasible to generate reliable and accurate synthetic CT scans from MR images with the proposed framework, opening up the possibility of harnessing the benefits of non-ionizing techniques to drive craniosynostosis diagnosis and treatment planning.