Prediction of fluid intelligence from T1-w MRI images: A precise two-step deep learning framework

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Predicting fluid intelligence from neuroimaging data is crucial for understanding cognitive development.
• The Adolescent Brain Cognitive Development (ABCD) Neurocognitive Prediction Challenge (ABCD-NP-Challenge) aims to advance predictive algorithms.
• Existing methods often struggle with accuracy in predicting fluid intelligence from T1-weighted MRI images.

Purpose of the Study:

• To propose a novel two-step deep learning pipeline to enhance the prediction accuracy of fluid intelligence scores from T1-weighted MRI data.
• To improve the segmentation of regions of interest (ROIs) in brain MRI scans.
• To achieve superior performance compared to existing state-of-the-art methods in fluid intelligence prediction.

Main Methods:

• A two-step deep learning pipeline integrating Residual Network (ResNet) and Squeeze-and-Excitation Network (SENet) concepts with a 3D U-Net architecture.
• A novel segmentation approach assigning identical labels to pixels in symmetrical brain regions.
• Utilizing a minimum bounding cube (MBC) for ROI segmentation and background information removal, followed by neural network regression models.

Main Results:

• The proposed segmentation structure achieved a Dice coefficient of 0.8920, significantly improving ROI segmentation performance over classical Convolutional Neural Networks (CNNs).
• The two-step deep learning pipeline demonstrated superior fluid intelligence score prediction accuracy.
• The method achieved a competitive Mean Square Error (MSE) of 82.56 on a test dataset.

Conclusions:

• The developed two-step deep learning pipeline effectively predicts fluid intelligence scores from T1-w MRI images.
• The integration of ResNet, SENet, and MBC techniques enhances both brain MRI segmentation and prediction accuracy.
• This approach represents a significant advancement in neurocognitive prediction, offering competitive performance in the ABCD-NP-Challenge.