Transfer learning using inception-ResNet-v2 model to the augmented neuroimages data for autism spectrum disorder classification

Abstract

From a psychiatric perspective, the detection of Autism Spectrum Disorders (ASD) can be seen from the differences in some parts of the brain. The availability of the four-dimensional resting-state Functional Magnetic Resonance Imaging (rs-fMRI) from Autism Brain Imaging Data Exchange I (ABIDE I) led us to reorganize it into two-dimensional data and extracted it further to create a pool of neuroimage dataset. This dataset was then augmented by shear transformation, brightness, and zoom adjustments. Resampling and normalization were also performed. Reflecting on prior studies, this classification accuracy of ASD using only 2D neuroimages should be improved. Hence, we proposed the use of transfer learning with the InceptionResNetV2 model on the augmented dataset. After freezing layer by layer, the best training, validation, and testing results were 70.22%, 57.75%, and 57.6%, respectively. We proved that the transfer learning approach was successfully outperformed the convolutional neural network (CNN) model from the previous study by up to 2.6%.