What question did this study set out to answer?

The aim is to assess the efficacy of synthetic FDG-PET images generated from MRI to aid in epilepsy evaluations.

January 18, 2026

Multisequence MRI Enables High‐Fidelity FDG ‐ PET Synthesis for Epilepsy Using GANs

Key Points

The aim is to assess the efficacy of synthetic FDG-PET images generated from MRI to aid in epilepsy evaluations.
Retrospective analysis of 481 participants with PET/MR
Divided into internal training (n=249), validation (n=31) and test sets (n=31)
Used SSIM, PSNR, Bland–Altman analysis for performance assessment
Visual quality grading by three blinded readers
Radiomics for classifying hippocampal sclerosis and outcome prediction
82.8% of visual ratings were excellent/good (κ = 0.42)
Internal SSIM was 0.98 ± 0.01 and external SSIM was 0.97 ± 0.01
SUVR correlation with ground-truth PET was r = 0.94 (internal) and r = 0.89 (external)
Detection accuracy for temporal hypometabolism was 90.3% (internal) and 87.1% (external)
Radiomics AUCs for detecting hippocampal sclerosis versus healthy controls was 0.72

Abstract

ABSTRACT Background FDG‐PET aids presurgical epilepsy evaluation but is limited by access and radiation exposure. Purpose To evaluate synthetic FDG‐PET generated from T1‐weighted imaging and resting‐state fMRI metrics. Study Type Retrospective. Population 481 participants underwent simultaneous FDG PET/MR. Internal cohort: 311 epilepsy patients split into training/validation/internal test sets ( n = 249/31/31; age 18.79 ± 16.33/22.20 ± 11.21/21.65 ± 17.62 years; male/female 145/104, 13/18, 22/9). External cohort: 115 temporal lobe epilepsy patients (age 25.36 ± 10.95 years; male/female 68/47) and 55 healthy controls (age 27.62 ± 5.82 years; male/female 24/31); 92 had surgery with 1‐year outcome. Field Strength Hybrid PET/MR at 3.0 T; gradient‐echo T1WI, echo‐planar imaging and resting‐state BOLD gradient‐echo EPI. Assessment Performance was assessed using SSIM, PSNR, MSE, NRMSE, SUVR correlation, and Bland–Altman analysis. Three blinded readers performed visual quality grading and detection of temporal lobe hypometabolism. Hippocampal radiomics was used for classification of hippocampal sclerosis and Engel outcome. Statistical Tests t ‐tests, chi‐square tests, Pearson correlation, Kolmogorov–Smirnov tests, DeLong tests, and false discovery rate correction. Results Excellent/Good visual ratings occurred in 82.8% (166/201), with Fleiss' κ = 0.42. SSIM was 0.98 ± 0.01 (internal) and 0.97 ± 0.01 (external); PSNR was 66.66 ± 1.25 and 64.16 ± 1.83, respectively. SUVR correlation with ground‐truth PET was r = 0.94 (internal) and r = 0.89 (external); Bland–Altman bias was −0.02 (95% limits of agreement: −0.22 to 0.18) internally and −0.00002 (95% limits: −0.35 to 0.35) externally. Detection accuracy for temporal hypometabolism was 90.3% (internal; κ = 0.735) and 87.1% (external; κ = 0.758). Radiomics AUCs using synthetic PET were 0.72 (95% CI: 0.62–0.83) for hippocampal sclerosis versus healthy controls and 0.77 (95% CI: 0.67–0.87) for Engel IA versus IB–IV; DeLong tests versus ground‐truth PET were non‐significant ( p = 0.56 and p = 0.48). Conclusion Multisequence MRI‐based synthetic PET showed high agreement with ground‐truth PET across image‐quality and quantitative SUVR metrics, providing a PET‐like metabolic surrogate when FDG‐PET is unavailable or impractical. Level of Evidence Evidence Level 3. Stages of Technical Efficacy Stage 3.

Mark Helpful

Bookmark

Relay