DCL-SE: Dynamic curriculum learning for spatiotemporal encoding of brain imaging

Area of Science:

• Neuroimaging analysis
• Computational neuroscience
• Medical diagnostics

Background:

• High-dimensional neuroimaging analyses face challenges with spatiotemporal fidelity and model adaptability.
• Existing large-scale, general-purpose models often compromise diagnostic accuracy and interpretability.

Purpose of the Study:

• Introduce Dynamic Curriculum Learning for Spatiotemporal Encoding (DCL-SE) for improved clinical diagnosis.
• Develop a framework for data-driven spatiotemporal encoding (DaSE) to enhance neuroimaging analysis.
• Address limitations in current neuroimaging analysis methods for clinical applications.

Main Methods:

• Utilize Approximate Rank Pooling (ARP) for efficient encoding of 3D volumetric brain data into 2D dynamic representations.
• Employ a Dynamic Graph Matching (DGM)-implemented architectural curriculum for progressive feature refinement.
• Implement a fixed hierarchical decoder that progresses from global anatomy to fine pathological details without stage-switching.

Main Results:

• DCL-SE consistently outperformed existing methods across six public datasets.
• Demonstrated superior accuracy, robustness, and interpretability in tasks including Alzheimer's disease and brain tumor classification, cerebral artery segmentation, and brain age prediction.
• Validated the effectiveness of compact, task-specific architectures in neuroimaging.

Conclusions:

• DCL-SE offers a significant advancement in high-dimensional neuroimaging analysis for clinical diagnosis.
• Task-specific architectures are crucial for overcoming limitations of large-scale pretrained networks in medical imaging.
• The DCL-SE framework provides a robust and interpretable solution for complex neuroimaging tasks.