Machine Learning–Based Synthetic Computed Tomography Derived From Temporal Bone Magnetic Resonance Imaging

Importance: For diagnostics and presurgical planning in otology, both magnetic resonance imaging (MRI) and computed tomography (CT) are frequently required to visualize soft tissues and bone structures. Ideally, visualization of both soft and bony tissues would be obtained using a single, radiation-free imaging modality. Objective: To train and evaluate a machine learning algorithm designed for generating synthetic CT images from MRI of the head. Design, Setting, and Participants: This diagnostic study assessing the feasibility of generating synthetic CT images from MRI of the head was conducted in a tertiary referral center in the Netherlands from September 2022 to September 2023. Paired MRI and CT scans were obtained from patients for whom CT of the head was requested as part of routine clinical care. Sixty-seven paired scans were used to train a machine learning algorithm to generate synthetic CT images from MRI data by a third party. Generated synthetic CT images of 15 patients were used for clinical evaluation by clinicians not associated with the third party. Main Outcomes and Measures: The primary outcomes were geometric and radiodensity accuracy, conspicuity of clinically relevant landmarks, and suitability of scans. Conspicuity on synthetic CT and true CT were independently rated by 2 ear, nose, and throat surgeons and 2 radiologists on a 4-point Likert scale ranging from poor (score of 1) to excellent (score of 4). Qualitative assessment of suitability of the scans for various clinical purposes was done by the 4 raters. Results: Paired MRI and CT scans were obtained from 73 patients (median age, 54 years [range, 18-81 years]; 38 [52%] male). Geometry and radiodensity on synthetic CT were sufficiently accurate (mean [SD] surface distance error, 0.38 [0.37] mm; mean [SD] radiodensity error, 4 [44] Hounsfield units) compared with true CT (gold standard). Conspicuity of the landmarks was generally deemed to be comparable. The thickness of the tegmen bone was sometimes overestimated by the algorithm, while the ossicles were often not depicted. Among 15 synthetic CT scans, each reviewed twice, most were considered suitable for localization (29 [97%]), navigation (25 [83%]), and surgical planning prior to cochlear implantation (21 [70%]) but not for diagnostic purposes. Conclusions and Relevance: In this study, CT-like images of the temporal bone could be generated from MRI of the head using a machine learning algorithm, which allowed for visualization of bony structures next to soft tissues using a single, radiation-free modality. The findings suggest that synthetic CT images are suitable for localization of anatomic structures in otologic procedures and reliable in estimating the extent of mastoid pneumatization for preoperative planning.