A classification method for autism based on multi-atlas transfer learning

Due to the lack of sufficient labeled samples and the selection of features based on a single brain map, it is difficult to fully explore the hidden features of different brain atlas in the current research on autism classification task, which leads to the problem of low accuracy of autism classification. For the first time, an innovative autism classification algorithm MATL based on multi-atlas transfer learning is proposed to extract and transfer features between different brain atlas, and to improve the deep neural network to achieve autism classification.The algorithm was trained on HO brain atlas and CC400 brain atlas using feature extractor, and the weight parameters of CC400 brain atlas feature extractor were obtained. Secondly, the weight was transferred to the HO brain atlas feature extraction network as the initial weight for training. Then, the output features were fed into an improved deep neural network classifier to complete the task of autism classification. The experiment was performed on ABIDE, a published functional magnetic resonance imaging dataset of autism, and the classification accuracy of the model reached 72.97%. The accuracy of MATL algorithm is 13.97 percentage points higher than that of non-transfer algorithm (wT-MATL). Compared with the baseline algorithm for autism classification task, the classification accuracy improved by 8.39 percentage points. The experimental results show that MATL algorithm can effectively fuse the data features of different brain atlas, and has better performance in autism classification tasks.