Validation of deep-learning-based MRI-to-CT attenuation correction for striatal and extrastriatal [¹²³I]I-FP-CIT...

Attenuation correction is critical for accurate SPECT brain imaging and to quantify uptake ratios in neurological and psychiatric disorders. This prospective study aimed to validate the use of deep-learning-based magnetic resonance to synthetic computed tomography (DL-MRAC) attenuation correction by quantifying striatal and extrastriatal binding of [¹²³I]I-FP-CIT to dopamine and serotonin transporters in Parkinson's disease patients.

Synthetic CTs were generated from T1-weighted MRIs acquired in 12 Parkinson's disease patients using a validated 3D residual U-Net for attenuation correction of [¹²³I]I-FP-CIT SPECT scans. We tested for equivalence of DL-MRAC versus CT-based attenuation correction (CTAC). We further compared uniform correction using Chang's method (UAC) and no attenuation correction (NAC) for regional analysis of specific binding ratios (SBRs) in striatal and extrastriatal areas. Data from the Parkinson's Progression Markers Initiative were used for external validation (n = 18).

As compared to CTAC, mean bias of DL-MRAC SBRs was -0.4% (95% confidence interval, CI -1.2 to 0.4) in the striatum and - 0.1% (95% CI, -0.8 to 0.6) in extrastriatal areas. UAC overestimated SBRs with mean bias ranging from 7.5% to 12.4%, whereas NAC underestimated SBRs with mean bias ranging from -6.5% to -24.0%, for striatal and extrastriatal binding estimates in all cohorts.

DL-MRAC is a valid, radiation-free method for attenuation correction and quantification of [¹²³I]I-FP-CIT binding to dopamine and serotonin transporters in subjects undergoing DaTSPECT examinations. It outperforms UAC and NAC and may therefore serve as a valuable alternative for patients for whom MRI is available and for data acquired on SPECT-only cameras.