Applying tensor-based morphometry to parametric surfaces can improve MRI-based disease diagnosis

Area of Science:

• Medical Imaging
• Computational Anatomy
• Machine Learning

Background:

• Computer-assisted diagnostic classification is crucial for identifying subtle differences in medical images.
• Existing tensor-based morphometry (TBM) methods have limitations in capturing complex anatomical variations.
• Machine learning with 3D brain maps offers potential for improved subject classification.

Purpose of the Study:

• To develop a novel approach for diagnostic classification using tensor-based morphometry on parametric surface models.
• To identify and utilize cortical surface features for enhanced classification accuracy.
• To compare the efficacy of the new method against existing TBM statistics.

Main Methods:

• Utilized holomorphic 1-forms for conformal mapping of multiply connected meshes.
• Employed surface parameterization with constrained harmonic maps for subject registration.
• Analyzed full Riemannian surface metric tensors and applied L1-norm based sparse learning for feature selection.
• Validated feature sets using stability selection.

Main Results:

• Multivariate statistics on local tensors showed greater effect sizes for detecting group differences compared to Jacobian determinant and eigenvalue analysis.
• The proposed method achieved reasonable classification results, outperforming Jacobian determinant, eigenvalue pairs, and volume features.
• Demonstrated effectiveness on MRI-derived cortical surfaces from Williams syndrome patients and controls.

Conclusions:

• The developed tensor-based morphometry approach on parametric surface models offers a powerful tool for diagnostic classification.
• This pipeline can enhance the sensitivity of morphometry studies and aid in image-based classification.
• The method shows promise for clinical applications in neurodevelopmental disorders.