Skull segmentation in 3D neonatal MRI using hybrid Hopfield Neural Network

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Accurate segmentation of neonatal skull in Magnetic Resonance (MR) images is crucial for source localization of electroencephalography (EEG) and magnetoencephalography (MEG) signals.
• Identifying signal sources helps pinpoint the origin of neurological abnormalities in neonates.
• Existing neonatal skull segmentation methods face challenges due to non-homogeneous skull intensities in MR images.

Purpose of the Study:

• To propose a fully automated hybrid algorithm for neonatal skull segmentation in MR images.
• To enhance the accuracy of source localization for EEG/MEG signals by improving skull segmentation.
• To address the challenge of non-homogeneous skull intensities in MR images.

Main Methods:

• A hybrid algorithm combining a Bayesian classifying framework with a Hopfield Neural Network (HNN) was developed.
• Local statistical parameters were utilized for adaptive training of the HNN, guided by Bayesian classifier error.
• The method was tested on high-resolution T1-weighted MR images from nine neonates (39-42 weeks gestational age).

Main Results:

• The proposed automated method achieved 65% accuracy in neonatal skull segmentation.
• The hybrid approach demonstrated superiority compared to previous neonatal skull segmentation techniques.
• Adaptive training based on local statistics and Bayesian error improved segmentation robustness.

Conclusions:

• The developed hybrid algorithm offers a robust and automated solution for neonatal skull segmentation.
• This method provides a foundation for more accurate source localization of EEG/MEG signals in neonates.
• The approach effectively handles intensity variations in MR images, improving segmentation performance.