Abstracts

Rationale: Temporal lobe epilepsy (TLE) is the most common type of epilepsy, and the Nucleus Basalis of Meynert (NBM) has recently emerged as one of the regions with the most disturbed connectivity [1]. Investigations into the NBM have been limited by the lack of robust segmentation strategies, instead relying on a probabilistic atlas [2]. Given that the NBM anatomy has been shown to change across disease states and that TLE shows increased gray matter atrophy, it is important to generate a patient specific segmentation of the NBM [3]. Patient specific segmentations are further complicated by the difficulty in accurately delineating it on 3T MRI, instead requiring 7T imaging (Figure 1). Deep learning approaches have shown success in segmenting other small, subcortical structures; thus, we investigated whether a deep learning network could accurately segment the NBM on commonly utilized 3T MRI.

Methods: To overcome the manual segmentation difficulty on 3T MRI, we obtained paired 3T and 7T MRI datasets of 21 healthy subjects and 14 patients with TLE. The NBM was manually segmented on 7T MRI with verification from a board-certified neurosurgeon. The 3T scans of each subject were rigidly registered to the 7T scans, providing accurate NBM segmentations with 3T MRI (Figure 1). Data augmentation was performed on the healthy subjects and 6 healthy subjects were withheld as a test dataset. A 3D-Unet convolutional neural network was constructed, and a 5-fold cross validation was performed for model selection. The model was tested on healthy subjects using the held-out test dataset and on the withheld dataset of 14 TLE patients.

Results: Qualitative comparison of the deep learning NBM segmentation revealed that the network could identify anatomical differences across a wide range of anatomical variability. Quantitative comparison of the deep learning and probabilistic atlas segmentation demonstrated significantly improved dice coefficient for both healthy subjects and patients with TLE. Additionally, the centroid distance was significantly decreased when using the deep learning network compared to the probabilistic atlas in patients with TLE (Figure 2).

Conclusions: We developed the first model, to our knowledge, for automatic and accurate patient specific segmentation of the NBM using deep learning. The deep learning network accurately identifies patient specific anatomical differences, performs better than the probabilistic atlas quantitatively, and captures the NBM in cases where the probabilistic atlas is unable to. This network may allow for further understanding of connectivity disturbances in TLE.

References:_x000D_ 1. Gonzalez HFJ, et al. Role of the nucleus basalis as a key network node in temporal lobe epilepsy. Neurology, 2021;96(9):e1334-e1346._x000D_ 2. Zaborszky L, et al. Stereotaxic probabilistic maps of the magnocellular cell groups in human basal forebrain. Neuroimage. 2008;42(3):1127-1141._x000D_ 3. Alvim MKM, et al. Progression of gray matter atrophy in seizure-free patients with temporal lobe epilepsy. Epilepsia. 2016;57(4):621-629.

Funding: This work was supported in part by NIH grants T32GM007347, T32EB001628-17, T32EB021937, F31NS106735, R01NS112252, R01NS108445, R01NS110130