Input Layer Regularization and Automated Regularization Hyperparameter Tuning for Myelin Water Estimation Using Deep...

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Machine Learning

Background:

• Myelin water fraction (MWF) is crucial for assessing brain tissue integrity.
• Accurate MWF estimation from magnetic resonance relaxometry is challenging.
• Biexponential signal modeling is a standard but complex method for MWF quantification.

Purpose of the Study:

• To develop a deep learning framework for improved MWF estimation.
• To integrate classical regularization and data preprocessing techniques.
• To enhance the accuracy of myelin content quantification in the brain.

Main Methods:

• A deep learning framework based on input layer regularization (ILR) was developed.
• Optimal regularization hyperparameters were selected using a neural network or generalized cross-validation (GCV).
• The framework was extended to directly estimate MWF and exponential time constants.

Main Results:

• The proposed deep learning architecture outperformed conventional methods on synthetic data.
• GCV-based hyperparameter selection showed slightly better performance than the neural network approach on in vivo data.
• The framework demonstrated superior accuracy in MWF estimation compared to standard multilayer perceptrons.

Conclusions:

• Input layer regularization significantly enhances MWF estimation within the biexponential model.
• Integrating classical regularization with deep learning substantially improves quantitative myelin content estimation.
• The developed framework offers a more accurate approach for brain myelin quantification using MRI data.