FAF-DRVFL: Fuzzy activation function based deep random vector functional links network for early diagnosis of Alzheimer disease

Abstract

Alzheimer’s disease (AD) is a degenerative neural condition marked by gradual memory loss and cognitive impairment. It is irreversible in nature and leads to progressive cerebral cortex atrophy. Therefore, structural Magnetic Resonance Imaging (sMRI) is an important tool that can be used for early-stage prediction of AD. Currently, deep learning networks are used for the diagnosis of AD, but it suffers from the limitations of gradient descent training of deep networks like local minima, slow learning speed, and overfitting. Also, there is a need to select hyperparameters like learning rate, momentum, number of epochs, and regularization coefficient. This paper proposes a deep non-iterative random vector functional link (RVFL) neural network. First, the MRI images’ features are extracted using transfer learning, and the classification of the extracted features is done using a non-iterative random vector initialized RVFL network. At the hidden layer of the RVFL classifier, the fuzzy activation function (FAF), is used to calculate the hidden layer’s output. The proposed algorithm has been evaluated and compared with the state-of-the-art methods on the ADNI dataset consisting of Cognitive Normal (CN), AD, converter Mild Cognitive Impairment (cMCI) and non-converter Mild Cognitive Impairment (ncMCI) MRI images. The performance achieved for CN vs AD diagnosis includes accuracy (86.67%), sensitivity (83.33%), specificity (88.89%), precision (83.33%), recall (83.33%) and F-score(86.07%) as well as Receiver Operating Characteristics shows that proposed method outperforms over several compared methods.