Machine learning-based method to detect capillary stalling using optical coherence tomography.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Biology

Background:

• Capillary stalling is a key factor in neurological disease models.
• Current quantification methods for capillary stalling are time-consuming and error-prone.
• Existing methods may not fully utilize 3D OCTA data.

Purpose of the Study:

• To develop an automated computational method for quantifying capillary stalling.
• To improve the accuracy and efficiency of capillary stall analysis.
• To investigate the morphological characteristics of stalled capillaries.

Main Methods:

• Developed a support vector machine (SVM) classifier using engineered features from time-series OCTA data.
• Validated the SVM classifier using 4-fold cross-validation.
• Compared the computational tool's performance against expert annotations.

Main Results:

• The SVM classifier achieved high performance with an ROC AUC of 0.978 and PR AUC of 0.700.
• Annotation time was reduced from 1 hour to 22 minutes per dataset.
• The tool identified 26% more stalling segments than expert annotations, revealing smaller diameter, increased tortuosity, and greater length in stalled segments.

Conclusions:

• The developed computational tool significantly enhances the efficiency and accuracy of capillary stall quantification.
• The tool facilitates the discovery of morphological patterns associated with capillary stalling.
• This open-source software supports further research into neurological disorders and capillary dynamics.