Article Details

Abstract

Autism Spectrum Disorder (ASD) is a common central neurodevelopmental disorder. Its clinical diagnosis faces problems of strong subjectivity and insufficient accuracy. To improve the performance of functional Magnetic Resonance Imaging (fMRI) in computer-aided diagnosis, this paper constructs a Lightweight Geometric Manifold Deep Network (LGMD-Net). This network adopts a multi-channel two-dimensional residual structure combined with medical geometric priors to obtain three-dimensional information, and achieves efficient feature extraction through lightweight residual blocks. In the feature dimension reduction stage, this paper introduces Spatial Pyramid Pooling to capture multi-scale spatial features, and combines Manifold Mixup technology to realize feature enhancement and regularization. This paper conducts experiments on the ABIDE dataset. The results show that this method outperforms Convolutional Neural Networks, Transformer, and Graph Neural Network models in terms of classification accuracy, F1 score, and AUC. Meanwhile, this method possesses significant advantages in parameter scale, inference time, and memory usage. The research results verify the feasibility of LGMD-Net in improving diagnostic accuracy and engineering application.

Keywords

ASD; fMRI; lightweight network; Spatial Pyramid Pooling; Manifold Mixup

References

1. [1] NEVISON C D. A Comparison of Temporal Trends in United States Autism Prevalence to Trends in Suspected Environmental Factors [J].Environmental Health，2014，13：73.
2. [2] HIROTA T，KING B H. Autism Spectrum Disorder: A Review [J].JAMA，2023，329（2）：157-168.
3. [3] MAKSIMOVIĆ S，MARISAVLJEVIĆ M，STANOJEVIĆ N，et al. Importance of Early Intervention in Reducing Autistic Symptoms and Speech-Language Deficits in Children with Autism Spectrum Disorder [J].Children，2023，10（1）：122.
4. [4] Yuan Yuzhuo, Luo Fang Screening and Diagnosis of Early Autism Patients Assisted by Artificial Intelligence [J]. Advances in Psychological Science, 2022, 30 (10): 2303-2320
5. [5] ZANG Y F，HE Y，ZHU C Z，et al. Altered Baseline Brain Activity in Children with ADHD Revealed by Resting-state Functional MRI [J].Brain and Development，2007，29（2）：83-91.
6. [6] DOBRYAKOVA E，HAFIZ R，IOSIPCHUK O，et al. ALFF Response Interaction with Learning During Feedback in Individuals with Multiple Sclerosis [J].Multiple Sclerosis and Related Disorders，2023，70：104510.
7. [7] JAHANI A，JAHANI I，KHADEM A，et al. Twinned Neuroimaging Analysis Contributes to Improving the Classification of Young People with Autism Spectrum Disorder [J].Scientific Reports，2024，14：20120.
8. [8] LI J L，ZHENG W H，FU X，et al. Individual Deviation-Based Functional Hypergraph for Identifying Subtypes of Autism Spectrum Disorder [J].Brain Sciences，2024，14（8）：738.
9. [9] Gu Jian, Qian Yurong, Wang Lanlan, etc A review of research on autism using artificial intelligence in functional magnetic resonance imaging data [J]. Computer Engineering and Applications, 2023, 59 (22): 57-68
10. [10] LECUN Y，BOSER B E，DENKER J S，et al. Handwritten Digit Recognition with a Back-Propagation Network [C]//Advances in Neural Information Processing Systems 2.Denver：Morgan Kaufmann，1990：396-404.
11. [11] THOMAS R M，GALLO S，CERLIANI L，et al. Classifying Autism Spectrum Disorder Using the Temporal Statistics of Resting-State Functional MRI Data with 3D Convolutional Neural Networks [J].Frontiers in Psychiatry，2020，11：440.
12. [12] JAHANI A，JAHANI I，KHADEM A，et al. Twinned Neuroimaging Analysis Contributes to Improving the Classification of Young People with Autism Spectrum Disorder [J].Scientific Reports，2024，14：20120.
13. [13] Li Zehui, Zhang Lin, Shan Xianying A review of the improvement and application of three-dimensional convolutional neural network methods [J]. Computer Engineering and Applications, 2025, 61 (3): 48-61
14. [14] FENG M，XU J. Detection of ASD Children Through Deep-Learning Application of fMRI [J].Children，2023，10（10）：1654.
15. [15] Liu Hongda, Sun Xuhui, Li Yibin, etc A review of deep learning models for image classification based on convolutional neural networks [J]. Computer Engineering and Applications, 2025, 61 (11): 1-21

About the authors

Wu Jinying (2000-), female, Han ethnicity, from Jining, Shandong, currently pursuing a master's degree in medical image processing; Corresponding author: Ma Huibin (1972-), female, Han ethnicity, from Jiamusi, Heilongjiang Province, professor, master's degree, research direction: machine learning and digital image processing.