Interpretable convolutional neural network for autism diagnosis support in children using structural magnetic...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Developmental Neuroscience

Background:

• Autism Spectrum Disorder (ASD) is a common neurodevelopmental condition with diverse symptoms, making early diagnosis challenging.
• Accurate and early diagnosis of autism is crucial for timely intervention and improved outcomes.
• Current diagnostic methods face challenges due to symptom variability and developmental changes in children.

Purpose of the Study:

• To evaluate the feasibility of an explainable deep learning (DL) model utilizing structural MRI (sMRI) for autism diagnosis in children.
• To identify significant brain biomarkers associated with autism through model interpretability.
• To support early and accurate diagnosis of autism in pediatric populations.

Main Methods:

• T1-weighted sMRI scans from 452 children (9-11 years) were analyzed from the Autism Brain Imaging Data Exchange.
• A DL model was developed to distinguish between autistic and typically developing children.
• Model explainability was assessed using saliency maps, and performance was compared to traditional machine learning models.

Main Results:

• The DL model achieved a mean area under the receiver operating curve of 71.2%.
• Saliency maps identified known autism-related neuroanatomical and functional biomarkers, including the cuneus, pericalcarine, ventricles, lingual and vermal lobules, caudate, and thalamus.
• The model demonstrated the ability to highlight key brain regions contributing to classification.

Conclusions:

• Interpretable DL models trained on sMRI data show promise in assisting autism diagnosis in a specific pediatric age group.
• These findings advance the application of explainable AI in neurodevelopmental research.
• The approach may contribute to improved clinical decision-making for autism and other neurodevelopmental disorders.