Early Detection of ASD Before Age 2 via Cortical Morphology and Explainable Geometric Deep Learning
Presented During:
Thursday, June 26, 2025: 11:30 AM - 12:45 PM
Brisbane Convention & Exhibition Centre
Room:
M1 & M2 (Mezzanine Level)
Poster No:
279
Submission Type:
Abstract Submission
Authors:
Qianyu Hou1,2, Chunfeng Lian1,2, Xianjun Li3, Jian Yang3, Fan Wang2,4, Jianhua Ma2,4
Institutions:
1School of Mathematics and Statistics, Xi'an Jiaotong University, Xi'an, China, 2Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University, Xi'an, China, 3Department of Radiology, The First Affiliated Hospital of Xi'an Jiaotong University, Xi'an, China, 4Key Laboratory of Biomedical Information Engineering of Ministry of Education, School of Life Science and Technology, Xi'an Jiaotong University, Xi'an, China
First Author:
Qianyu Hou
School of Mathematics and Statistics, Xi'an Jiaotong University|Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
School of Mathematics and Statistics, Xi'an Jiaotong University|Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
Co-Author(s):
Chunfeng Lian
School of Mathematics and Statistics, Xi'an Jiaotong University|Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
School of Mathematics and Statistics, Xi'an Jiaotong University|Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
Xianjun Li
Department of Radiology, The First Affiliated Hospital of Xi'an Jiaotong University
Xi'an, China
Department of Radiology, The First Affiliated Hospital of Xi'an Jiaotong University
Xi'an, China
Jian Yang
Department of Radiology, The First Affiliated Hospital of Xi'an Jiaotong University
Xi'an, China
Department of Radiology, The First Affiliated Hospital of Xi'an Jiaotong University
Xi'an, China
Fan Wang
Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University|Key Laboratory of Biomedical Information Engineering of Ministry of Education, School of Life Science and Technology, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University|Key Laboratory of Biomedical Information Engineering of Ministry of Education, School of Life Science and Technology, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
Jianhua Ma
Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University|Key Laboratory of Biomedical Information Engineering of Ministry of Education, School of Life Science and Technology, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
Research Center for Intelligent Medical Equipment and Devices, Xi'an Jiaotong University|Key Laboratory of Biomedical Information Engineering of Ministry of Education, School of Life Science and Technology, Xi'an Jiaotong University
Xi'an, China|Xi'an, China
Introduction:
Autism Spectrum Disorder (ASD) is typically diagnosed through behavioral assessments at ages 2 to 3 or later. However, Hazlett et al. demonstrated the feasibility of early diagnosis by identifying significant brain overgrowth between 6 and 24 months, which can serve as an early indicator of ASD (Hazlett, 2017). While neuroimaging studies offer promising insights, rapid structural changes during the first two years and the heterogeneous neurological abnormalities in ASD hinder stable and consistent diagnosis. To address these issues, we propose NeuroExplainer, a novel surface-based analysis approach for more precise and interpretable early ASD diagnoses.
Methods:
This study utilizes the IBIS dataset (Gilmore, 2012), which is publicly available and comprises 76 autism and 278 healthy cases. Cortical thickness, surface area, and volume measurements were computed at 6, 12, and 24 months, following the protocol in (Yin, 2024). Linear data augmentation was applied to increase the training set to 1,206 cases, while the test set remained at 73. Confounding variables were eliminated through linear regression analysis following (Wang, 2022).
As shown in Fig. 1, a self-attention operation is initially applied to the concatenation of the encoder output feature maps to capture inter-hemispheric dependencies. Then, both class predictions and an attention map are derived by the Spherical Attention block, which consists of global average pooling and 1D convolution. The complete decoder is a hierarchical structure that employs transpose convolutions and skip connections. This allows the network to generate fine-grained attention maps, facilitating precise pathological localization.
Besides, to improve capturing interpretable factors, we design regularization terms for fidelity and sparsity according to (Yuan, 2022). Specifically, the attention maps produced by our NeuroExplainer, denote as \( A_i^+ \) and \( A_j^- \) respectively, should be sparse. And, we assume that the feature vector for autism cases \( f_i^+ = 1^T(A_i^+ 1_{1 \times C} \odot F_i^+) \) highlights autism-relevant areas, and the rest of the cerebral cortex \( \bar{f}_i^- = 1^T((1 - A_i^+)1_{1 \times C} \odot F_i^+) \) still grows normally. To the end, we design losses as
\[
L_{sparsity} = \sum_{i \neq j} 1^T \{ A_i^+ \odot \log(A_i^+) - A_j^- \odot \log(A_j^-) \}
\]
\[
L_{fidelity} = \sum_{i \neq j} \| \bar{f}_i^+ - f_j^- \| + \max(m - \| f_j^- - f_i^+ \|, 0)
\]
As shown in Fig. 1, a self-attention operation is initially applied to the concatenation of the encoder output feature maps to capture inter-hemispheric dependencies. Then, both class predictions and an attention map are derived by the Spherical Attention block, which consists of global average pooling and 1D convolution. The complete decoder is a hierarchical structure that employs transpose convolutions and skip connections. This allows the network to generate fine-grained attention maps, facilitating precise pathological localization.
Besides, to improve capturing interpretable factors, we design regularization terms for fidelity and sparsity according to (Yuan, 2022). Specifically, the attention maps produced by our NeuroExplainer, denote as \( A_i^+ \) and \( A_j^- \) respectively, should be sparse. And, we assume that the feature vector for autism cases \( f_i^+ = 1^T(A_i^+ 1_{1 \times C} \odot F_i^+) \) highlights autism-relevant areas, and the rest of the cerebral cortex \( \bar{f}_i^- = 1^T((1 - A_i^+)1_{1 \times C} \odot F_i^+) \) still grows normally. To the end, we design losses as
\[
L_{sparsity} = \sum_{i \neq j} 1^T \{ A_i^+ \odot \log(A_i^+) - A_j^- \odot \log(A_j^-) \}
\]
\[
L_{fidelity} = \sum_{i \neq j} \| \bar{f}_i^+ - f_j^- \| + \max(m - \| f_j^- - f_i^+ \|, 0)
\]
·Fig. 1. The framwork of our NeuroExplainer architecture and Spherical attention block.
Results:
Figure 2(a) highlight the competitive performance of NeuroExplainer across multiple metrics. While UGSCNN achieves the highest accuracy with fewer parameters, NeuroExplainer outperforms in all metrics, especially in sensitivity (0.79), which is crucial for clinical applications. Moreover, it significantly decreases the parameters compared to SiT, balancing model complexity and performance. Despite class imbalance, which reduces the sensitivity of other methods, NeuroExplainer achieves the highest sensitivity and distinguishes features associated with autism.
Figure 2(b) illustrates attention maps from different methods on cortical surfaces. Chebnet, Spherical UNet and SiT display dispersed patterns, while UGSCNN and MoNet tend to highlight partial cortical regions. In contrast, NeuroExplainer delivers more precise and comprehensive attention localization. Specifically, ASD related cortical areas are identified as left superior frontal gyrus, anterior cingulate cortex, temporal regions (STS, STG), fusiform gyrus, and right orbitofrontal cortex. These regions align with (Yin, 2024), demonstrating the meaningfulness of our method in capturing fine-grained autism-related developmental features.
Figure 2(b) illustrates attention maps from different methods on cortical surfaces. Chebnet, Spherical UNet and SiT display dispersed patterns, while UGSCNN and MoNet tend to highlight partial cortical regions. In contrast, NeuroExplainer delivers more precise and comprehensive attention localization. Specifically, ASD related cortical areas are identified as left superior frontal gyrus, anterior cingulate cortex, temporal regions (STS, STG), fusiform gyrus, and right orbitofrontal cortex. These regions align with (Yin, 2024), demonstrating the meaningfulness of our method in capturing fine-grained autism-related developmental features.
·Fig. 2. (a) Classification results and (b)Visualization comparison of attention maps on cortical surfaces across different geometric deep learning models.
Conclusions:
In this paper, we propose NeuroExplainer, a method designed to learn fine-grained interpretable features of Autism Spectrum Disorder (ASD) on the cortical surface. Our approach demonstrated superior classification performance and identifies cortical areas associated with ASD, offering valuable insights into its neural underpinnings.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Methods Development
Keywords:
Autism
Computational Neuroscience
Cortex
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
Provide references using APA citation style.
[2]Hazlett, H. C., Gu, H., Munsell, B. C., Kim, S. H., Styner, M., Wolff, J. J., ... & Statistical Analysis Gu Core H. 17. (2017). Early brain development in infants at high risk for autism spectrum disorder. Nature, 542(7641), 348-351.
[3]Wang, Y., Hu, D., Wu, Z., Wang, L., Huang, W., & Li, G. (2022). Developmental abnormalities of structural covariance networks of cortical thickness and surface area in autistic infants within the first 2 years. Cerebral Cortex, 32(17), 3786-3798.
[4]Xue, C., Wang, F., Zhu, Y., Li, H., Meng, D., Shen, D., & Lian, C. (2023, October). Neuro Explainer: Fine-Grained Attention Decoding to Uncover Cortical Development Patterns of Preterm Infants. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 202-211). Cham: Springer Nature Switzerland.
[5]Yin, Z., Ding, X., Zhang, X., Wu, Z., Wang, L., Xu, X., & Li, G. (2024). Early autism diagnosis based on path signature and Siamese unsupervised feature compressor. Cerebral Cortex, 34(13), 72-83.
[6]Yuan, H., Yu, H., Gui, S., & Ji, S. (2022). Explainability in graph neural networks: A taxonomic survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5), 5782-5799.
[7]Zhao, F., Xia, S., Wu, Z., Duan, D., Wang, L., Lin, W., ... & Li, G. (2019). Spherical U-Net on cortical surfaces: Methods and applications. In Information Processing in Medical Imaging: 26th International Conference, IPMI 2019, Hong Kong, China, June 2–7, 2019, Proceedings 26 (pp. 855-866). Springer International Publishing.