PCA Based Hierarchical CNN for the Classification of Mild Cognitive Impairments and the Role of SIREN Activations

The use of Convolutional Neural Networks for the classification of volumetric data is contentious because 2-D convolutions miss out on the correlation between the slices of the volume, whilst 3-D networks guzzle extensive computing resources. Moreover, the advent of SIREN activations calls for the investigation regarding the role of activations in such networks. This work proposes a model that uses the Principal Component Analysis to reduce the given data, followed by a circumspectly designed CNN for extracting the pertinent features. The paper also investigates the role of activations in such networks. The method is used to classify the patients converted to Alzheimer's from Mild Cognitive Impairment from those who did not convert. The data is obtained from ADNI. The proposed work gives an accuracy of 94.29, which is better as compared to the state-of-the-art.