Objective: The purpose of this study was to evaluate the utility of global semi-quantitative analysis via fluorine-18-flurodeoxyglucose positron emission tomography (18F-FDG PET) at lateralizing seizure foci and diagnosing patients with unilateral temporal lobe epilepsy (TLE).
Methods: Seventeen patients with unilateral TLE (11 right TLE and 6 left TLE) were retrospectively selected for semi-quantitative 18F-FDG PET analysis. Twenty-three control subjects with a Mini Mental State Examination (MMSE) score of 29 or greater were selected for comparison. Globally averaged standardized uptake value (gSUVmean) was computed for each temporal lobe. Lateralization indices (LI) and the absolute value of lateralization indices (|LI|) were calculated to assess the degree of asymmetry in each subject. Logistic regression analyses were performed at a probability cutoff of 0.5 to classify TLE patients as left or right TLE and to discriminate patients from control subjects. Receiver operating characteristic (ROC) curves were generated to evaluate the utility of LI and |LI| as classification predictors. The Bland Altman test was used to evaluate the reproducibility of the measurements.
Results: There was a statistically significant difference in gSUVmean computed LI between left and right TLE patients (P<0.01). There was no statistically significant difference in |LI| between the patient and control groups (P=0.22). Logistic regression revealed that 82% of TLE patients were lateralized correctly using LI as the sole predictor. The area under the ROC curve (AUC) was 0.80. Logistic regression using |LI| on the combined patient/control population showed a diagnostic accuracy of 65% and an AUC of 0.44. Bland Altman analysis revealed an intra-observer reproducibility of 96% and an inter-observer reproducibility of 96% and 91% on successive trials.
Conclusion: We conclude that gSUVmean computed LI is a reliable and reproducible measure for predicting seizure lateralization in unilateral TLE patients. However, gSUVmean computed |LI| does not appear to be particularly effective at diagnosing TLE patients from control subjects. Further studies with more patients should investigate other machine learning techniques that combine gSUVmean with other diagnostic predictors.