Optimize Transfer Learning for Autism Spectrum Disorder Classification with Neuroimaging: A Comparative Study
Lakmini Herath, Dulani Meedeniya, Janaka Marasingha, Vajira S. Weerasinghe
Abstract
The field of neuroimage classification using deep neural networks (DNN) is a fast-moving area. With the absence of sufficient neuroimaging data, transfer learning based DNN plays a significant contribution in image classifications. The associated optimizers support generating promising results using these deep learning models. This study explores the optimal combinations of hyperparameters to achieve the high classification accuracy of fMRI data to distinguish Autism Spectrum Disorder (ASD) subjects from the control group. The proposed Inception-v3 based DNN was trained with transfer learning in four different modes with generated 2D images from fMRI images. The optimizers Adaptive moment estimation (Adam) and Stochastic Gradient Descent (SGD) with a learning rate of 1E-04 and 1E-03, respectively, yield the highest average accuracy value of 98.38% and 98.06%.