Multiple-Activation Parallel Convolution Network in Combination with t-SNE for the Classification of Mild Cognitive Impairment

The classification of Mild Cognitive Impairment can be done using 2-D CNN, which take a single slice at a time as input and do not consider pixel information from adjacent slices or spatial correlation amongst the slices of the brain volume or 3-D CNN, which requires huge computation time and memory as a significantly large number of parameters involved in 3D-CNN in comparison to 2D-CNN. To reduce the spatial correlation, computational complexity, and memory requirement, we use t-Distributed Stochastic Neighbor Embedding (t-SNE) on MRI volume to reduce its dimensions. Also, we use parallel CNN instead of sequential to analyze MRI volumes and a combination of RELU, sigmoid, and SIREN activation functions to learn better features for the classification of MCI. To check the efficacy of the proposed t-SNE Multiple-Activation Parallel Convolution Network, experiments are performed on publicly available Alzheimer's Disease Neuroimaging Initiative dataset, and performance is compared with existing methods. We obtain classification accuracy of 94.15 and 94.89 on MCI-C Vs. MCI-NC data and MCI Vs. Controls data respectively.