Brain-Oct-Pvt: A Physics-Guided Transformer with Radial Prior and Deformable Alignment for Neurovascular...

Area of Science:

• Medical Imaging
• Deep Learning
• Neuroscience

Background:

• Neurovascular Optical Coherence Tomography (OCT) imaging presents unique challenges for deep learning segmentation.
• Existing models like Polyp-PVT have limitations in handling grayscale data, vascular curvature, radial patterns, and thin boundaries.

Purpose of the Study:

• To develop a specialized deep learning framework, Brain-OCT-PVT, tailored for neurovascular OCT imaging.
• To overcome the limitations of current models in accurately segmenting neurovascular structures.

Main Methods:

• Introduced a single-channel input stem to reduce computational load.
• Developed a Radial Intensity Module (RIM) for annular structure analysis using polar transforms.
• Implemented a Deformable Cross-scale Fusion Module (D-CFM) and Boundary-aware Attention Module (BAM) for improved feature fusion and edge detection.
• Utilized a specialized loss function combining Dice Similarity Coefficient, BoundaryIoU, and Focal Tversky for class imbalance.

Main Results:

• Achieved a Dice score of 95.06% on clinical neurovascular OCT data.
• Obtained a 95% Hausdorff Distance (HD95) of 0.269 mm.
• Demonstrated superior performance compared to existing segmentation approaches.

Conclusions:

• Brain-OCT-PVT effectively addresses the challenges in neurovascular OCT segmentation.
• The proposed framework shows significant potential for clinical applications in analyzing brain vasculature.