Leveraging Input-Level Feature Deformation With Guided-Attention for Sulcal Labeling.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Cortical sulci identification is crucial for understanding brain development and function.
• While primary and secondary sulci are well-studied, tertiary sulci remain under-investigated due to anatomical variability and data scarcity.
• Automatic labeling of cortical sulci presents challenges including high variability, small region sizes, and limited annotated data.

Purpose of the Study:

• To develop a novel end-to-end learning framework for accurate automatic labeling of cortical sulci, with a focus on challenging tertiary sulci.
• To address the limitations of existing methods in handling anatomical variability and small regions of interest.
• To improve the understanding of functional and structural brain development through enhanced sulci identification.

Main Methods:

• Proposed a spherical convolutional neural network (CNN) framework for end-to-end learning.
• Developed a novel feature warping technique to mitigate anatomical variability during sulci labeling.
• Introduced a guided-attention mechanism to focus on discriminative sulcal regions and suppress irrelevant information.

Main Results:

• The proposed method demonstrated superior performance in automatic cortical sulci labeling compared to existing approaches.
• Significant improvements were observed particularly in the accurate identification of putative tertiary sulci.
• The framework effectively handles anatomical variability and emphasizes relevant sulcal regions.

Conclusions:

• The novel deep learning framework offers a robust solution for automatic cortical sulci labeling, especially for variable tertiary sulci.
• The method's ability to manage anatomical variability and focus on key regions advances neuroimaging analysis.
• This work provides a valuable tool for researchers studying cortical development and function.