| In recent years,autism has an increasing impact on the healthy growth of children and adolescents.Due to the difficulty of diagnosis and treatment of autism in the later stage,it is extremely important to conduct advanced diagnosis and early intervention and treatment in the early stage,especially in the infant period.With the development of MRI technology,effective treatment of autism through the brain’s MRI data is constantly being explored.At present,MRI still has many challenges for diagnosing infants with autism:(1)The data acquisition is difficult and the amount of data is small;(2)The number of data categories is unbalanced;(3)The data feature categories are not rich enough;(4)The heterogeneity of the brains of different individuals.Based on the above challenges,the main research content and contributions of this paper are as follows:Aiming at the problem of the small amount of data of autism MRI data,this paper proposes a twin similarity learning network,which transforms the classification idea,converts singlesample input into two-sample paired input,and increases the data input volume of the network model.The mathematical tool of path signature is introduced to extract and supplement the information in the time dimension of the infant brain magnetic resonance data.The model structure after the combination of the model and the feature is a very effective method for the diagnosis of autism.Aiming at the problem of autism and the unbalanced number of normal samples in infant structural magnetic resonance data,an autism classification network based on dual-channel autoencoder and twin similarity learning network combined with learning is proposed.In order to reduce the learning difficulty caused by data imbalance,after feature expansion,the network structure of a dual-channel autoencoder is proposed to perform unsupervised feature compression.In order to eliminate the ambiguity when learning similar paired samples,multitask constraint learning is proposed on the basis of the twin similarity network.At the same time,in the test classification stage,due to the heterogeneity between different samples,the similarity weighted voting method is used to obtain the category similarity of the test data.In summary,this paper studies the diagnosis of infant autism in three aspects: feature extraction,model structure and heterogeneity constraints.In terms of feature extraction,the path reintegration feature is introduced to extract effective brain development features from the time dimension;in terms of model structure,the model structure of dual-channel autoencoder and twin similarity learning network is proposed to improve the model’s resistance to imbalances.Data processing ability;in terms of heterogeneity constraints,a reasonable similarity voting strategy based on the similarity of features of different brain regions is proposed in the testing phase,which reduces the impact of noise data.Finally,this article carried out the ablation experiment of each module on the NDAR autism database and the comparison experiment with other algorithms.The experimental results prove that our method has very excellent performance for the classification and recognition of autism.For the full amount of NDAR data,we can achieve a 65% autism recall rate and an overall classification accuracy rate of 87.7%. |
