SEPoolConvNeXt: A Deep Learning Framework for Automated Classification of Neonatal Brain Development Using T1- and...

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Neonatal and infant brain development involves rapid myelination and cortical maturation, critical for assessing normative trajectories and detecting abnormalities.
• Automated classification of brain development using MRI is challenging due to overlapping developmental stages and sex-specific variations.

Purpose of the Study:

• To introduce SEPoolConvNeXt, a novel deep learning framework for fine-grained classification of neonatal brain development.
• To evaluate the performance of SEPoolConvNeXt against established Convolutional Neural Network (CNN) models using T1- and T2-weighted MRI sequences.

Main Methods:

• Developed SEPoolConvNeXt, integrating residual connections, grouped convolutions, and channel attention for efficient and discriminative analysis.
• Utilized a dataset of 29,516 MRI images across four subgroups (T1 Male, T1 Female, T2 Male, T2 Female), stratified into 14 age classes (0-12 months).
• Compared SEPoolConvNeXt performance against 19 pre-trained CNNs under identical experimental conditions.

Main Results:

• SEPoolConvNeXt achieved test accuracies consistently above 95%, significantly outperforming baseline CNNs (average ~70.7%).
• High accuracies were observed across all subgroups, with T2 Female reaching 99.47%-100% and T1 Male exceeding 98%.
• The model demonstrated robust generalization, with most subgroups exceeding 98-99% accuracy in combined evaluations.

Conclusions:

• SEPoolConvNeXt offers a robust, efficient, and biologically relevant framework for assessing neonatal brain maturation.
• The model's ability to integrate sex- and age-specific developmental trajectories provides a foundation for AI-assisted neurodevelopmental evaluation.
• This approach shows promise for clinical translation, particularly for monitoring high-risk infants like preterm babies.