Hybrid optimization enabled DenseNet for autism spectrum disorders using MRI image.

Abstract

Autism spectrum disorder (ASD) is the neuro-developmental disorder caused by various changes in the brain. It affects the life conditions with social interaction and communication. Most of the previous researches used the various techniques for the early detection to reduce the ASD, but it had been occurred several complications such as, time expenses, and low accessibility for diagnosis.This paper aims to develop the JSTO-DenseNetmodel is for the detection of ASD. In this paper, an input autism brainimage is considered as an input applied to image pre-processing phase. In image pre-processing, the clatters are removed utilizing Gaussian filtering and also, Region of Interest (ROI) extraction is carried out. Thereafter, extraction of pivotal region is done based on functional connectivity utilizing proposed Jaya Sewing Training Optimization (JSTO). The JSTO is newly introduced by combining Jaya algorithm and Sewing Training-Based Optimization (STBO). Thus, output-1 is obtained. In feature extraction phase, grey level co-occurrence matrix (GLCM) features like entropy, correlation, energy, homogeneity, inverse difference moment, Angular second moment and texture features namelylocal ternary patterns (LTP), Local Optimal Oriented Pattern (LOOP) and Histogram of Oriented Gradients (HOG) are extracted from the Magnetic Resonance Imaging (MRI). Therefore, output-2 is obtained. From output-1 and output-2, ASD classification is accomplished using DenseNet, which is trained employing same proposed JSTO.The proposed JSTO-DenseNet model achieves the highest accuracy of 94.8 %, True Positive Rate (TPR) of 90 %, True Negative Rate (TNR) of 90.5 %, un-weighted average recall (UAR) of 89.8 % and the lowest False Negative Rate (FNR) of 86.7 %, and False Positive Rate of 82.6 %, when compared with other traditional methods like, Explainable Artificial Intelligence (XAI), Hybrid deep lightweight feature generator, CLAttention, Two stream end-to-end deep learning (DL), Auto-Encoder feature representation, and Fuzzy Inference Gait System-Deep Extreme Adaptive Fuzzy (FIGS-DEAF) based on Abide 1 dataset.