Deep learning segmentation of the nucleus basalis of Meynert on 3T MRI

Abstract

The nucleus basalis of Meynert (NBM) is a key subcortical structure that is important in arousal, cognition, brain network modulation, and has been explored as a deep brain stimulation target. It has also been implicated in several disease states, including Alzheimers disease, Parkinsons disease, and temporal lobe epilepsy (TLE). Given the small size of NBM and variability between patients, NBM is difficult to study; thus, accurate, patient-specific segmentation is needed. We investigated whether a deep learning network could produce accurate, patient-specific segmentations of NBM on commonly utilized 3T MRI. It is difficult to accurately segment NBM on 3T MRI, with 7T being preferred. Paired 3T and 7T MRI datasets of 21 healthy subjects were obtained, with 6 completely withheld for testing. NBM was expertly segmented on 7T MRI, providing accurate labels for the paired 3T MRI. An external dataset of 14 patients with TLE was used to test the model on brains with neurological disorders. A 3D-Unet convolutional neural network was constructed, and a 5-fold cross-validation was performed. The model was evaluated on healthy subjects using the held-out test dataset and the external dataset of TLE patients. The model demonstrated significantly improved dice coefficient over the standard probabilistic atlas for both healthy subjects (0.68MEAN{+/-}0.08SD vs. 0.47{+/-}0.06, p=0.0089, t-test) and TLE patients (0.63{+/-}0.08 vs. 0.38{+/-}0.19, p=0.0001). Additionally, the centroid distance was significantly decreased when using the model in patients with TLE (1.22{+/-}0.33mm, 3.25{+/-}2.57mm, p=0.0110). We developed the first model, to our knowledge, for automatic and accurate patient-specific segmentation of the NBM.