FaithfulNet: An explainable deep learning framework for autism diagnosis using structural MRI.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning models offer potential for diagnosing complex neurological conditions like autism but often function as "black boxes."
• Explainable Artificial Intelligence (XAI) techniques are crucial for interpreting these models, increasing trust in clinical applications.
• Understanding the neural correlates of autism is vital for developing effective interventions.

Purpose of the Study:

• To develop and validate an explainable deep learning model for autism diagnosis using sMRI data.
• To identify specific brain regions associated with autism and their impact on academic performance through XAI.
• To create a novel, faithful visual explanation method (Faith_CAM) for deep learning predictions in autism.

Main Methods:

• Utilized the ABIDE-II dataset of structural Magnetic Resonance Imaging (sMRI) data.
• Developed FaithfulNet, a deep learning model for autism diagnosis.
• Applied gradient-based class activation maps and SHAP gradient explainer for model interpretability.
• Integrated explanations to create Faith_CAM, quantified using the pointing game score and analyzed with brain structure masks.

Main Results:

• Achieved a high classification accuracy of 99.74% and an Area Under the Curve (AUC) of 1 for autism diagnosis.
• Successfully identified impaired brain regions in individuals with autism.
• Quantified the impact of these impairments on memory regions linked to academic performance.

Conclusions:

• The developed XAI approach, Faith_CAM, provides a trustworthy and interpretable method for autism diagnosis.
• This study successfully links autism diagnosis with specific neural impairments affecting cognitive functions and academic performance.
• Findings support the development of personalized treatment strategies for children with autism.