Connectional-style-guided contextual representation learning for brain disease diagnosis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning (DL) in brain disease diagnosis using structural magnetic resonance imaging (sMRI) often focuses on local features, leading to poor generalization.
• Current methods may capture spurious information, limiting their effectiveness across different diseases and datasets.
• Understanding intrinsic brain patterns beyond single domains is crucial for robust diagnostic models.

Purpose of the Study:

• To develop a novel deep learning model for multi-disease brain diagnosis that captures intrinsic brain patterns.
• To enhance the generalization ability and robustness of diagnostic models by focusing on connectional properties.
• To improve the accuracy of computer-aided diagnosis for brain diseases.

Main Methods:

• Proposed a connectional style contextual representation learning model (CS-CRL) utilizing a vision transformer (ViT) encoder.
• Employed mask reconstruction as a proxy task and Gram matrices to guide the representation of connectional information.
• Focused on capturing global context and biologically plausible feature aggregation.

Main Results:

• CS-CRL achieved superior accuracy in diagnosing multiple brain diseases across six datasets and three diseases.
• The model outperformed existing state-of-the-art methods in diagnostic performance.
• Demonstrated that CS-CRL captures brain-network-like properties, aggregates features effectively, is easier to optimize, and is more robust to noise.

Conclusions:

• CS-CRL offers a more robust and accurate approach to computer-aided brain disease diagnosis compared to previous methods.
• Capturing connectional brain properties provides a more biologically relevant and generalizable feature representation.
• The model's ability to learn global context and aggregate features enhances its clinical utility for diverse neurological conditions.