Abstracts

Rationale: We have recently shown using group-wise comparison of diffusion kurtosis imaging (DKI) data, that temporal lobe epilepsy (TLE) patients are characterized by microstructural changes along white matter (WM) bundles connected to the temporal pole (Kasa et al. 2021). DKI allows quantification of diffusion in complex tissue environments and is potentially more sensitive to diffusion anomalies (Jensen et al. 2005). Nonetheless, it remains unclear if these observed changes are restricted to the temporopolar area and/or could be detected automatically at the patient level. This work aims to assess this using Detect, a recently proposed anomaly detection framework based on an autoencode (AE) network (Chamberland et al., 2021): Twenty-five TLE patients and 24 healthy controls (HC) were recruited for this study. Lesional TLE patients (n=15) showed mesial temporal abnormalities, while absent for non-lesional TLE patients (n=10). A T1-weighted MPRAGE image (1 mm3) and two 2 mm3 diffusion-weighted scans (b=0, 1300, 2600 s/mm, 130 diffusion directions, with reverse phase-encoding, respectively) were acquired for each subject using a 3T MRI system. PyDesigner was used for DWI data processing (Dhiman et al. 2021) while mean kurtosis (MK) maps were obtained using the Diffusion Kurtosis Estimator tool (v2.6). For detecting microstructural anomalies along WM bundles, we trained Detect with 80% of the HC data and a 20% validation set composed of equal numbers of patients and HC. To test for loss, 10% of the training set is held out during the training. Detect aims to generate output similar to the input by minimizing the mean absolute error (MAE), the difference between raw and reconstructed WM-tract DKI measurements, serving as the ‘anomaly score’. We compared the performance of the AE network with z-scoring and principal component analysis (PCA). All results were corrected for age and sex differences.

Results: The AE approach showed better discriminative power compared to PCA and z-scoring in the lesional group (Fig. 1). High anomaly scores were seen in the left inferior fronto-occipital fasciculus (IFO) and right inferior longitudinal fasciculus (ILF). For an example patient, anomalies were detected along the left IFO and right ILF,  supporting the neuroradiological findings of left hippocampus and right temporal lobe alterations in this patient. No clear differences were seen in the performance of the three approaches for the non-lesional patients (Fig. 2). However, high anomaly scores in left IFO agree with our individual patient analysis and intracranial EEG characterization.

Conclusions: Combining DKI and AE appears more successful in detecting local anomalies along WM-tracts than PCA and z-scoring. Increased variability in the WM tract features in the non-lesional group might underlie the lower performance in this group. Identifying local, patient-specific abnormalities along the tracts in TLE could improve outcomes and help move towards personalized treatment.

Funding: CIHR Foundation, NSERC Discovery, Canada First Research Excellence Fund, Brain Canada, EpLink