Indentification of autism spectrum disorder based on an improved convolutional neural networks

In order to develop the application of deep learning for identification of autism with small sample size, an improved algorithm based convolutional neural networks (CNN) was proposed in this paper. Compared to the classical CNN, the 2-dimensional input was changed to a 1-dimensional input, PReLU was adopted as the activation function instead of ReLU, and the softmax was used as classifier at the time of pre-training and replaced with the parameter-adjustable SVM after pre-training. Furthermore, L2 regularization was used for cross entropy loss function in order to avoid overfitting. A total of 416 subjects with autism and 441 subjects with typical development across 14 sites were involved in the study. The original feature was extracted from the resting-state functional magnetic resonance imaging data. The improved CNN could identify the autism with an accuracy of 84.44% (sensitivity 85.39%, specificity 80.57%) by the 10-fold cross-validation. The results demonstrated that the proposed algorithm performs better than those in recent similar studies.