Informative Feature-Guided Siamese Network for Early Diagnosis of Autism

• Published in: Machine Learning in Medical Imaging Vol. 12436; pp. 674 - 682
• Main Authors: Gao, Kun, Sun, Yue, Niu, Sijie, Wang, Li
• Format: Book Chapter Journal Article
• Language: English
• Published: Switzerland Springer International Publishing AG 01.01.2020; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Autism spectrum disorder; Early diagnosis; Siamese network; Subject-specific attention; Siamese network; Early diagnosis; Subject-specific attention; Autism spectrum disorder
• ISBN: 9783030598600; 3030598608
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-59861-7_68

Autism, or autism spectrum disorder (ASD), is a complex developmental disability, and usually diagnosed with observations at around 3-4 years old based on behaviors. Studies have indicated that the early treatment, especially during early brain development in the first two years of life, can significantly improve the symptoms, therefore, it is important to identify ASD as early as possible. Most previous works employed imaging-based biomarkers for the early diagnosis of ASD. However, they only focused on extracting features from the intensity images, ignoring the more informative guidance from segmentation and parcellation maps. Moreover, since the number of autistic subjects is always much smaller than that of normal subjects, this class-imbalance issue makes the ASD diagnosis more challenging. In this work, we propose an end-to-end informative feature-guided Siamese network for the early ASD diagnosis. Specifically, besides T1w and T2w images, the discriminative features from segmentation and parcellation maps are also employed to train the model. To alleviate the class-imbalance issue, the Siamese network is utilized to effectively learn what makes the pair of inputs belong to the same class or different classes. Furthermore, the subject-specific attention module is incorporated to identify the ASD-related regions in an end-to-end fully automatic learning manner. Both ablation study and comparisons demonstrate the effectiveness of the proposed method, achieving an overall accuracy of 85.4%, sensitivity of 80.8%, and specificity of 86.7%.