Relating occlusion maps obtained through deep learning to functional impairment in dementia of Alzheimer’s type: Neuroimaging/Optimal neuroimaging measures for early detection

Jinhyeong Bae, Jane Stocks, Ashley Heywood, Youngmoon Jung, Aggelos Katsaggelos, Lisanne M Jenkins, Karteek Popuri, Mirza Faisal Beg, Lei Wang

Abstract:

Background
Predicting the conversion from Mild Cognitive Impairment (MCI) into Dementia of the Alzheimer’s type (DAT) and functional change is crucial to patient care and treatment. In order to visualize brain regions which are significant in the prediction, we implemented an occlusion map based on deep learning.
Method
Using T1‐weighted structural MRI data from ADNI, 3D convolutional neural network was trained to predict the conversion from MCI to DAT through a transfer learning pipeline. The model resulted in an 82.4% classification accuracy on an independent test set. An occlusion map was subsequently generated as follows. Each brain scan was occluded by 2x2x2 voxel patch iterated through every position in the brain. The model produced a prediction score corresponding to the location of the occlusion patch to identify important voxels for the model’s prediction at the subject level. Mean intensity …