Abstract
242253
Introduction: Mesial temporal lobe (MTLE), a common intractable epilepsy, often entails hippocampal damage. Surgery fails to stop seizures in many cases due to complex brain network injuries. 18F-FDG-PET scans are sensitive to metabolic changes in the epileptogenic zone (EZ) and can show hypometabolism beyond the EZ, sometimes extending to other brain areas. While diffusion MRI tracks white matter integrity, revealing disruptions that may cause persistent seizures. The link between white matter alterations and metabolic dysfunction is not yet clear.
This study employs graph neural network (GNN), a novel artificial intelligence network, to explore the relationship between structural connectivity and FDG PET uptake.
Methods: Data acquisition and processing
Data from 51 drug refractory unilateral MTLE patients (age > 18) were obtained on a 3T PET/MR scanner. PET was acquired around 45 minutes post 18F-FDG injection. Simultaneous MR data include: T1-weighted MPRAGE, and diffusion MRI. Other clinical records, including semiology, and radiological diagnoses of MR-HS and MR-negative cases were also collected.
Standardized uptake value ratios (SUVR) for 74 cortical and subcortical regions on the PET images were calculated and normalized using cerebellar gray. Diffusion images underwent motion and eddy current correction, followed by QSDR reconstruction using DSI-studio. Deterministic tractography was performed. Fractional anisotropy (FA), number of tracts between each pair of regions and mean diffusivity (MD) for each region were calculated.
Graph Network and evaluation
Graph nodes, representing specific brain regions, with features including MR-HS or MR-negative label, ipsilateral or contralateral hemisphere, age, age at epilepsy onset, illness duration, seizure frequency, region volume (normalized by estimated total intracranial volume), and mean MD. Edge weights between nodes were determined by FA values and the number of tracts for the corresponding pair. Self-loop was added. The network structure is shown in Figure 1. The network's task is to predict the SUVR of each region using individual’s graph. Analysis focused on regions known to be associated with MTLE and seizures, as shown in Figure 2a.
The data from 51 patients were split into a training and testing set (41:10). Results below are based on testing set only.
Results: Figure 2a presents the Spearman correlation results. Notably, transverse temporal cortex, middle temporal cortex, superior temporal cortex, insula, putamen, and pallidum exhibit high correlation on both sides. Additionally, caudate, parahippocampal, inferior temporal cortex, and fusiform gyrus show significant correlation on one side. Figure 2b contains scatter plots depicting the network predicted SUVR against measured SUVR for representative regions.
Conclusions: The study reveals that a GNN effectively correlates Diffusion MRI-based structural connectivity with regional FDG PET uptake in key areas associated with MTLE. Notably, temporal regions, including transverse temporal, middle, and superior temporal regions, show an expected high correlation due to the typical impact of MTLE on temporal lobes. Subcortical regions like the caudate and putamen exhibit significant correlation, indicating their connectivity might reflect broader MTLE effects beyond the temporal lobe. However, lower correlation in limbic regions (parahippocampal, hippocampus, amygdala) suggests their connectivity changes have less predictive power. Additionally, inferior temporal, thalamus, and associative areas (fusiform, entorhinal) show lower correlation, implying less direct correlation with metabolic activity despite potential MTLE influence. Predicting regional FDG PET uptake using diffusion data holds promise for non-invasive epileptogenic zone characterization crucial for surgical planning.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.