Dyslexia detection using 3D convolutional neural networks and functional magnetic resonance imaging

Area of Science:

• Neuroscience
• Developmental Psychology
• Machine Learning

Background:

• Dyslexia is a neurological disorder impacting reading and writing skills, often leading to significant academic challenges if not identified early.
• Early detection and intervention are vital for children with dyslexia to foster positive self-esteem and maximize academic potential.
• Undiagnosed dyslexia can cause frustration and achievement gaps, underscoring the need for effective diagnostic tools.

Purpose of the Study:

• To present a novel approach for the automatic recognition of dyslexia in children.
• To leverage functional magnetic resonance imaging (fMRI) data for identifying brain activation patterns associated with dyslexia.
• To develop a deep learning model for enhanced dyslexia detection.

Main Methods:

• fMRI scans were preprocessed, including motion correction, normalization, and smoothing, to enable voxel-wise comparisons.
• Statistical Parametric Maps (SPMs) were used to create 165 3D brain activation volumes from 55 children.
• A deep learning architecture, comprising three parallel 3D Convolutional Neural Networks (3D CNNs), was employed for classification, utilizing a 4-fold cross-validation strategy.

Main Results:

• The developed system achieved an overall average classification accuracy of 72.73% for dyslexia detection.
• The model demonstrated a sensitivity of 75% and a specificity of 71.43%.
• Precision was recorded at 60%, with an F1-score of 67%.

Conclusions:

• The study confirms the feasibility of using deep learning combined with fMRI for recognizing dyslexia in children.
• Functional magnetic resonance imaging during phonological and orthographic reading tasks provides valuable data for dyslexia identification.
• The proposed system offers a promising avenue for early and accurate dyslexia detection.