Gaussian smoothing and modified histogram normalization methods to improve neural-biomarker interpretations for...

Area of Science:

• Neuroimaging and computational anatomy.
• Development of advanced image processing techniques for biomedical applications.

Background:

• Interpreting neural-biomarkers for dyslexia from MRI data is challenging, especially with multi-source, heterogeneous datasets.
• Inconsistent scanner settings across different MRI data sources complicate biomarker analysis and interpretation.

Purpose of the Study:

• To present a novel method for improving the biological interpretability of dyslexia neural-biomarkers from diverse MRI datasets.
• To enhance the quality and comparability of neuroimaging features for dyslexia research using advanced image processing.

Main Methods:

• Utilized a modified histogram normalization (MHN) method combined with Gaussian smoothing (isotropic kernel, 4mm) to normalize MRI image intensities.
• Employed image resizing, labeling, patching, and non-rigid registration as pre-processing steps.
• Evaluated performance through ROI segmentation, gray matter volume estimation, and deep learning classification using CAT12 and pre-trained models.

Main Results:

• Optimal Gaussian smoothing achieved at σ = 1.25, yielding a 0.9% PSNR increase.
• The MHN and Gaussian smoothing methods significantly improved feature comparability, evidenced by high DSC and low MSE.
• Deep learning models achieved a maximum accuracy of 94.7% (95% CI) after 10x 10-fold cross-validation.

Conclusions:

• The proposed MHN method significantly outperforms state-of-the-art histogram matching for normalizing MRI data in dyslexia studies.
• The combined MHN and Gaussian smoothing approach enhances the reliability of neural-biomarkers and features for dyslexia research.
• This method offers a robust solution for analyzing heterogeneous MRI datasets, improving diagnostic accuracy and biological interpretation.