| Autism Spectrum Disorder(ASD)is a complex neurodevelopmental disorder that affects communication,socialization,and behavior.It is a lifelong condition that usually starts in early childhood and affects an individual’s entire life.With appropriate intervention and support,individuals with ASD can recover to live normal lives,making timely diagnosis crucial for treatment and recovery of normal development.The traditional diagnostic method is based on observation of behavioral symptoms,which can be prone to misdiagnosis due to the influence of children’s expression ability and doctor’s subjective judgment.Therefore,a more effective,automated,and reproducible detection method is needed to aid in the diagnosis of ASD.With the development of deep learning techniques and medical imaging technology,researchers have explored the use of brain structural information obtained from functional magnetic resonance imaging(f MRI)for diagnosing Autism Spectrum Disorder(ASD).However,due to the complexity of ASD lesions,training robust ASD classification models using deep learning networks remains challenging.In ASD diagnosis,low-level semantic information plays a crucial role in distinguishing between normal and abnormal brain regions.High-level features often capture more abstract and discriminative information,while low-level features capture local patterns and details,providing fine-grained information.By combining high-level and low-level features,the model can benefit from their complementary information,helping to better differentiate ASD patients from the normal population.This article focuses on the extraction and utilization of low-level information and designs three multi-scale networks based on three backbone networks.By establishing the fusion of high and low-level information,the network performance is improved,and the model accuracy is enhanced.The main content is as follows:(1)Based on 3D-ResNet,we propose a classification model that combines a multiscale information fusion mechanism with an attention subnet.The model selects an appropriate fusion strategy to comprehensively utilize the output of different levels of the network to fully capture information at different levels.The introduction of an attention mechanism can strengthen the extraction of lesion region features by the network,thereby improving model accuracy.(2)Res2Net is an extension of ResNet that aims to address the limitations of ResNet in multiscale feature modeling and aggregation.Based on 3D-Res2Net,we design an ASD classification model that can select the appropriate fusion strategy for multiscale information at different layers by introducing a context enhancement module.In addition,we also design a residual attention module,which can help the model focus on the important features of f MRI data and avoid being disturbed by irrelevant information.(3)Swin Transformer is a backbone network based on Transformer that can avoid some of the limitations of CNN,such as insufficient global information and missing long-term dependence information.We first construct a powerful baseline based on Swin Transformer,and then design two modules to enhance the ability to extract robust features from f MRI data.The ST context enhancement module aggregates multiscale feature information and enhances feature discrimination,while the cross-block module rearranges block sequences through displacement and shuffling operations to improve the ability to capture subtle local features.The models are trained and tested on the Autism Brain Imaging Data Exchange(ABIDE)dataset.Experimental results show that the three methods proposed in this paper achieve better classification performance compared to existing ASD classification models. |
