Deep Relation Learning for Regression and Its Application to Brain Age Estimation.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Medical Imaging

Background:

• Deep learning models for temporal regression often overlook inter-image relationships.
• Accurate brain age estimation is crucial for understanding neurological development and aging.

Purpose of the Study:

• To propose a novel deep relation learning framework for temporal regression tasks.
• To enhance brain age estimation by learning diverse non-linear relationships between image pairs.

Main Methods:

• Developed a deep neural network with feature extraction (CNN) and relation regression (Transformer) components.
• Simultaneously learned four non-linear relations: cumulative, relative, maximal, and minimal.
• Utilized an efficient convolutional neural network for deep feature extraction.

Main Results:

• Achieved a mean absolute error (MAE) of 2.38 years in brain age estimation on a large dataset (6,049 subjects).
• Demonstrated statistically significant improvement over 8 state-of-the-art algorithms (p<0.05).
• The proposed method effectively leverages relationships between input images.

Conclusions:

• Deep relation learning offers a superior approach for temporal regression tasks like brain age estimation.
• The simultaneous learning of multiple non-linear relations enhances prediction accuracy.
• This framework holds promise for advancing neuroimaging analysis and age-related research.