AI Model Reads Gray and White Matter Signals to Spot Autism with 87% Accuracy

Autism spectrum disorder has long been one of the most challenging conditions to diagnose objectively. Clinicians still rely largely on behavioral observation, which can vary widely between specialists and often delays identification of a condition where early intervention matters enormously. Now a team of researchers in China has combined a new generation of artificial intelligence architecture with brain imaging to detect autism with an accuracy of 87.49 percent, while simultaneously revealing previously unreported patterns in how the brain’s gray matter and white matter communicate with one another. The study, published in BMC Medical Imaging, analyzed resting-state functional magnetic resonance imaging and structural magnetic resonance imaging data from 1,175 participants drawn from the multi-center ABIDE I and ABIDE II datasets, one of the largest publicly available collections of autism neuroimaging data in the world.

The research, led by Xuan Yu and Meiyun Wang of the Medical Imaging Research Institute at Henan Provincial People’s Hospital and Zhengzhou University, set out to solve a persistent technical problem. Brain activity is not static. Functional connectivity, the statistical relationship between activity in different brain regions, fluctuates from moment to moment during a scanning session, and these dynamic patterns are thought to carry crucial information about neurodevelopmental disorders. Traditional analytical approaches, including many earlier deep learning models, struggle to capture both the spatial organization of these networks and their temporal evolution at the same time. Convolutional neural networks excel at spatial pattern recognition but have limited ability to model long-range dependencies in time, while recurrent architectures can track sequences but are computationally expensive and prone to losing information over long spans.

The team’s solution was a hybrid framework built on the Mamba architecture, a relatively new class of model known as a selective state space model. State space models process sequences by maintaining an internal state that is updated as each new data point arrives, and the selective variant learns to decide which parts of the input stream are relevant at each step. This allows the model to filter out noise and focus on the most informative dynamic interactions, while handling long-range temporal dependencies with far greater efficiency than conventional recurrent networks. In the framework described in the study, the Mamba component performs a two-dimensional selective scan, referred to as SS2D, which lets the network capture dependencies across both spatial and temporal dimensions of the imaging data simultaneously.

Working alongside the Mamba backbone, a convolutional neural network extracts local spatial features from the brain images, and the two streams are brought together through what the authors call a Hierarchical Multi-scale Fusion module, or HMSF. This module integrates brain network features across different spatiotemporal scales, from fine-grained local fluctuations to broader, slower patterns of coordination between distant regions. The fused representation is then passed through multi-layer perceptron layers for the final classification, with the model trained using binary cross entropy loss. The design reflects a growing recognition in computational neuroscience that meaningful biomarkers of disorders like autism do not live at a single scale but emerge from interactions across multiple levels of brain organization.

What distinguishes this study from earlier machine learning efforts is not just the classification performance but the way the researchers interrogated the trained model to extract biological insight. For the first time, they systematically characterized the specific features of dynamic network connectivity between gray matter, where neuronal cell bodies reside and information processing occurs, and white matter, the insulated fiber tracts that carry signals between regions, in people with autism. This gray matter to white matter perspective is relatively unexplored, as most connectivity studies focus exclusively on relationships between gray matter regions while treating white matter as passive wiring rather than an active partner in network dynamics.

The analysis revealed distinctive abnormal patterns in the autism group. Among them, the researchers identified enhanced connectivity between the superior longitudinal fasciculus, a major white matter tract that links frontal, parietal and temporal regions and supports language and attention, and the dorsal anterior cingulate cortex, a gray matter area involved in cognitive control and error monitoring. Conversely, they found reduced connectivity between the corpus callosum, the largest commissural tract connecting the two cerebral hemispheres, and the somatosensory association cortex, which integrates touch and bodily sensations. These findings fit a broader picture emerging from autism research of atypical long-range connectivity, with some pathways over-communicating and others under-communicating relative to neurotypical brains, and they point to specific tract-cortex pairs that could serve as imaging biomarkers for future diagnostic work.

On the classification task, the framework achieved 87.49 percent accuracy, 87.31 percent sensitivity and 87.68 percent specificity in distinguishing individuals with autism from healthy controls, outperforming other ASD classification methods tested in the study, including more conventional approaches such as support vector machines. The balance between sensitivity and specificity matters in a clinical context: a model that flags too many healthy individuals creates false alarms and unnecessary anxiety, while one that misses true cases undermines the entire purpose of screening. An area under the receiver operating characteristic curve analysis further supported the model’s discriminative ability. The authors report that the model was trained and evaluated with standard preprocessing safeguards, including control of head motion as measured by framewise displacement, a common confound in functional imaging studies of children and clinical populations.

The scale and diversity of the data strengthen the result. By pooling 1,175 participants across the ABIDE I and ABIDE II consortia, the researchers mitigated the site-specific artifacts that plague single-center neuroimaging studies, where a model may learn scanner characteristics rather than biology. Multi-center validation is widely regarded as an essential step toward clinically deployable AI tools, and the fact that the framework held up across heterogeneous acquisition parameters suggests the learned features reflect genuine neural signatures rather than equipment quirks. The study was conducted in accordance with the Declaration of Helsinki, with ethics approval from the Research Committee of Henan Provincial People’s Hospital and written informed consent obtained from a parent or guardian of each child participant, along with assent from the children themselves.

Beyond the immediate diagnostic application, the work contributes to a theoretical shift in how scientists understand autism’s neural mechanisms. Rather than viewing the disorder as a problem confined to specific gray matter regions, the findings support a network-level view in which the dynamic dialogue between processing centers and their connecting fibers is fundamentally altered. The enhanced coupling between the superior longitudinal fasciculus and the dorsal anterior cingulate cortex, for example, may relate to the atypical attentional control and sensory processing often observed in autism, while diminished corpus callosum involvement echoes decades of structural and functional evidence pointing to reduced interhemispheric communication in the condition. Linking these patterns to measurable dynamic connectivity opens the door to using them as targets for monitoring the effects of intervention over time.

The technical contribution may prove equally influential. Mamba-based state space models are rapidly gaining traction across machine learning for their ability to handle long sequences with linear computational cost, and this study demonstrates their value in medical imaging, where both spatial detail and temporal dynamics must be captured at scale. The Hierarchical Multi-scale Fusion strategy offers a template that other researchers could adapt to other disorders and imaging modalities, from Alzheimer’s disease to schizophrenia, wherever the interplay of space and time in brain networks holds diagnostic clues. The research was supported by the Science and Technology Project of Henan Province, the National Natural Science Foundation of China, the National Key R&D Program of China and the Joint Fund of Henan Province Science and Technology R&D Program, and the authors declare no competing interests. As open-access work, the study invites replication and extension by groups worldwide, a crucial step before any AI diagnostic tool can move from the journal page to the clinic.

Subject of Research: Gray and white matter functional connectivity patterns in autism spectrum disorder analyzed with a Mamba-CNN deep learning framework

Article Title: Study on spatio-temporal patterns of gray matter and white matter functional connectivity in autism spectrum disorder based on Mamba-CNN framework

Article References: Yu, X., Wu, Y., Bai, Y., Jin, S., Si, Y., Tan, J., Meng, N., Kong, Y., Zhang, X., Lin, Y., Xin, H., & Wang, M. (2026). Study on spatio-temporal patterns of gray matter and white matter functional connectivity in autism spectrum disorder based on Mamba-CNN framework. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02798-y

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02798-y

Keywords: autism spectrum disorder, functional connectivity, gray matter, white matter, Mamba, convolutional neural network, state space models, fMRI, ABIDE, deep learning, neuroimaging, BMC Medical Imaging