A Deep Learning Framework for Classifying Autism Spectrum Disorder from fMRI Images

The objective of this work is to propose a new machine learning based approach for ASD classification from fMRI data employing deep learning architecture and novel preprocessing method. Main datasets included in ABIDE I, ABIDE II, Human Connectome Project (HCP), and Social Functions and Autism Research Interface (SFARI). Many supplements include graph neural networks for brain segmentation, 3D-BRA for accurate mapping of the brain than just using conventional registration technique, EN-CC for feature selection. The approach incorporated graph neural network for segmentation of the brain, 3DBRA mapping that translates and rotates the brain precisely unlike the general technique of registration, EN-CC for selecting the features. The deep learning models under consideration were ResNet-50, VGG-16, Inception-v3, and Efficient Net which were improved through ABSC and MSTC. Classification accuracy of the proposed models was higher than 95% on all the datasets with Efficient Net scoring 99% on the SFARI dataset. This means that only the Stratified Group Cross-Validation (SGCV) was employed to make sure the results would be more reliable. The results present a considerable improvement in both accuracy and the interpretability of the derived connectivity profiles of the ASD-related brain.

Hali tarjima qilinmagan