Modified Exigent Features Block in JAN Net for Analysing SPECT Scan Images to Diagnose Early-Stage Parkinson's Disease.

Abstract

The quantitative measure of dopamine transporter (DaT) in the human midbrain is generally used as a biomarker for analyzing Parkinson's disease (PD). DaT scan images or Single- photon emission computed tomography (SPECT) images are utilized to capture the dopamine content more accurately. Only sixteen slices out of ninety-one of SPECT images were chosen on the basis of the high amount of dopamine content and were named Volume rendering image slices (VRIS). This paper proposes a novel Convolutional Neural Network (CNN) called JAN Net which particularly treats the VRIS for identifying PD. The JAN Net preserves the edges and spatial features of the striatum by using a modified exigent feature (M-ExFeat) block, that contains convolutional and additive layer. The different-sized convolutional layer extracts both low- and high-level features of Striatum. The additive layer adds up all the features of different filter sized convolutional layers like 1x1, 3x3, and 5x5. The added output features are used to improve the learnability of neurons in the hidden layer. The network performance is tested for stride 1 and stride 2. The results are validated using the dataset taken from the Parkinson's Progression Markers Initiative (PPMI) database. The JAN Net ensures improved performance in terms of accuracy. The training and validation accuracy for stride 2 is 100% with minimum losses. The outcome has been compared with different deep learning architectures and the machine learning techniques like Extreme Learning Machines (ELM), and Artificial Neural Networks (ANN) to highlight the efficacy of the proposed architecture. Hence, the present work could be of great aid to the experts in neurology to protect the neurons from impairment.