cGAN-Based Synthetic CT for MRI-Guided Transcranial Focused Ultrasound: A Feasibility Study

Abstract

MRI-guided transcranial focused ultrasound (tFUS) is a promising therapeutic technique for treating various neurological and psychiatric disorders. For accurate targeting of acoustic focus through the skull, however, an additional scan of CT is required. In this study, we demonstrated the generation of synthetic CT (sCT) from T1- weighted MRI using a deep learning method and investigated its feasibility in acoustic simulation. MRI and CT from 15 subjects were used to train the developed 3D conditional generative adversarial network (3D-cGAN). We estimated the quality of sCT from 10 independent test subjects in terms of skull characteristic and acoustic simulation.The skull density ratio (SDR) and skull thickness (ST) derived from sCT were compared with those of real CT (rCT). The acoustic simulation was performed under the same sonication conditions for both CTs. Consequently, 1500 sets of simulation were conducted (i.e., 50 transducer locations × 3 targets × 10 subjects). The k-Wave simulation toolbox was employed using a 200kHz single-element transducer. For skull characteristic, high correlations were shown between sCT and rCT (r = 0.92, p < 0.001 for SDR; r = 0.96, p < 0.001 for ST). No significant difference was found for both factors. The acoustic simulation results also showed high similarity. The intracranial peak acoustic pressure ratio was under 4%. The dice coefficient similarity of acoustic focus was about 0.85 and the distance between focal points was less than 1mm. The result of acoustic simulation suggests the applicability of deep learning-based sCT for MRI-guided tFUS in the future.