Multi-center brain age prediction via dual-modality fusion convolutional network

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Accurate brain age prediction is vital for understanding neurodevelopment and disease progression.
• Existing models struggle with multi-center datasets, especially those with small sample sizes.
• Current brain age prediction often treats it as a regression or classification problem, limiting accuracy.

Purpose of the Study:

• To develop a robust multi-center data correction method for brain age prediction.
• To introduce a dual-modality fused convolutional neural network (BrainDCN) for improved prediction accuracy.
• To enhance the generalizability and clinical applicability of brain age prediction models on small-sample datasets.

Main Methods:

• A multi-center data correction strategy using Wasserstein distance and maximum mean discrepancy.
• A brain dual-modality fused convolutional neural network (BrainDCN) with a joint loss function (MAE and cross-entropy).
• Feature weighting using matrices and vectors derived from single-center training sets for multi-center application.

Main Results:

• The BrainDCN model achieved the lowest average absolute error on the CamCAN dataset compared to state-of-the-art models.
• The multi-center correction method demonstrated superior performance across four neuroimaging datasets, outperforming existing methods.
• Application to multi-center schizophrenia data revealed mean accelerated aging compared to controls.

Conclusions:

• The proposed multi-center correction method and BrainDCN model significantly improve brain age prediction generalizability and accuracy.
• The joint loss function and weighted features further enhance prediction performance.
• This research provides a foundational methodology for multi-center brain age prediction, with strong clinical potential for small-sample studies.