Longitudinally Consistent Individualized Prediction of Infant Cortical Morphological Development

Area of Science:

• Neuroscience
• Developmental Biology
• Artificial Intelligence

Background:

• Infant neurodevelopment is critical, with many disorders originating from abnormal brain development.
• Existing methods for predicting neurodevelopmental trajectories from incomplete data have limitations in flexibility and consistency.
• Current deep learning models often predict missing data independently, ignoring longitudinal dependencies.

Purpose of the Study:

• To develop a novel framework for predicting individualized longitudinal cortical developmental trajectories from incomplete infant neurodevelopmental data.
• To address the limitations of existing regression and deep learning models in handling incomplete longitudinal data.
• To improve the understanding of normal early brain development and aid in the identification of neurodevelopmental disorders.

Main Methods:

• Developed a longitudinally consistent triplet disentanglement autoencoder (LCTDA) framework.
• Employed a surfaced-based autoencoder to disentangle identity-related and age-related features.
• Utilized dynamic time-warping loss to ensure temporal consistency and similarity among predicted trajectories.

Main Results:

• The LCTDA framework demonstrated superior longitudinal consistency and exactness in predicting infant cortical property maps.
• The model successfully generated individualized longitudinal developmental trajectories.
• Experimental results validated the framework's effectiveness compared to existing baseline methods.

Conclusions:

• The proposed LCTDA framework offers a flexible and temporally consistent approach for predicting neurodevelopmental trajectories from incomplete data.
• This method enhances the ability to study normal early brain development and detect neurodevelopmental disorders.
• The disentanglement of identity and age features allows for more accurate and individualized predictions.