Synthetic computed tomography of the head from magnetic resonance imaging based on latent diffusion model.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing 100053, China (W.L., T.S., Y.Z., Q.Z., J.L.); Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Xuanwu Hospital, Beijing 100053, China (W.L., T.S., Y.Z., Q.Z., J.L.); Key Laboratory of Neurodegenerative Diseases, Ministry of Education, Beijing 100053, China (W.L., T.S., Y.Z., Q.Z., J.L.).
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing 100053, China (W.L., T.S., Y.Z., Q.Z., J.L.); Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Xuanwu Hospital, Beijing 100053, China (W.L., T.S., Y.Z., Q.Z., J.L.); Key Laboratory of Neurodegenerative Diseases, Ministry of Education, Beijing 100053, China (W.L., T.S., Y.Z., Q.Z., J.L.). Electronic address: [email protected].
Abstract
Magnetic resonance imaging (MRI) is widely used in head scans. However, MRI lacks electron density information, which is inherent to computed tomography (CT). This study aimed to synthesize head CT from MRI using latent diffusion model (LDM). We collected 466 paired head T1-weighted/ultrashort echo time (UTE) MRI and CT images. The deep learning framework employed finite scalar quantization-generative adversarial network (FSQ-GAN) to compress CT images into latent representations. T1-weighted (one channel) or UTE MRI (three channels) were integrated through conditional embeddings within the LDM to guide CT synthesis. Comprehensive evaluations were conducted on the synthetic CT (sCT), including quality assessment, consistency of CT value distributions, and agreement of compositions between synthetic and real CT. sCT images generated by the FSQ-GAN and LDM framework achieved a mean absolute error of around 24 HU, a structural similarity index greater than 0.90, and a peak signal-to-noise ratio of 29 dB. The Pearson's r between CT values in the sCT and real CT images was 0.96, the Bhattacharyya distance, Kullback-Leibler divergence and Jensen-Shannon divergence between sCT and real CT distributions were 0.08, 0.70 and 0.24, respectively. Across different compositions, sCT yielded Dice coefficients exceeding 0.85 and 95th percentile Hausdorff distances under 24.79 mm relative to real CT. Furthermore, the sCT based on UTE MRI exhibited better quality than those generated from T1-weighted MRI. The proposed deep learning framework was capable of generating high-quality CT images from T1-weighted or UTE MRI, providing a robust methodological foundation for potential downstream clinical applications.