Deep Granular Feature-Label Distribution Learning for Neuroimaging-based Infant Age Prediction

Area of Science:

• Neuroimaging and Machine Learning
• Developmental Neuroscience

Background:

• Accurate infant age prediction from neuroimaging is vital for monitoring brain development.
• Existing methods often struggle with insufficient data, limiting their clinical applicability.
• Label distribution learning (LDL) is a machine learning approach suitable for small sample problems.

Purpose of the Study:

• To develop an accurate infant age prediction model using neuroimaging data, addressing data scarcity.
• To introduce and evaluate a novel deep granular feature-label distribution learning (DGFLDL) model.

Main Methods:

• Proposed granular label distribution (GLD) to assemble adjacent age labels into granules, reducing label count and augmenting data.
• Introduced granular feature distribution (GFD) to leverage image variability for enhanced learning effectiveness.
• Integrated GLD and GFD with deep neural networks, forming the DGFLDL model, using 8 cortical morphometric features from 384 infant MRI scans.

Main Results:

• The DGFLDL model achieved a mean absolute error of 36.1 days in infant age prediction.
• Demonstrated significant performance improvement over baseline methods.
• Feature importance analysis identified key biomarkers for infant brain development.

Conclusions:

• DGFLDL effectively addresses data scarcity in neuroimaging-based infant age prediction.
• The model offers a promising tool for analyzing infant brain development and identifying critical biomarkers.
• This approach advances the application of machine learning in developmental neuroscience.