Deep learning-based multiclass segmentation in aneurysmal subarachnoid hemorrhage.

Area of Science:

• Neurosurgery
• Radiology
• Artificial Intelligence

Background:

• Traditional radiological scoring for subarachnoid hemorrhage (SAH) suffers from variability and incomplete data utilization.
• Image segmentation offers precise delineation and potential for automated assessment of SAH extent.
• Developing automated tools is crucial for improving diagnostic accuracy and patient outcome prediction in SAH.

Purpose of the Study:

• To develop a deep learning model for automated multiclass segmentation of pathologies associated with aneurysmal subarachnoid hemorrhage.
• To assess the model's performance against human raters and in external validation datasets.
• To explore the potential of automated segmentation for creating imaging biomarkers for SAH outcome prediction.

Main Methods:

• Utilized 73 non-contrast CT scans from aneurysmal subarachnoid hemorrhage patients.
• Manually segmented six classes: subarachnoid, intraventricular, intracerebral, and subdural hemorrhage, aneurysms, and ventricles.
• Employed the nnU-Net deep learning framework (2D and 3D configurations) and performed interrater reliability analysis and external validation.

Main Results:

• The nnU-Net model achieved segmentation performance comparable to senior raters for key hemorrhage and ventricle classes.
• Median Dice coefficients for hemorrhage segmentation were 0.664 (3D) and 0.673 (2D) in the internal test set.
• External validation on primary intracerebral hemorrhage patients yielded a median Dice coefficient of 0.831 for hemorrhage segmentation.

Conclusions:

• Deep learning facilitates automated multiclass segmentation of SAH-related pathologies with near-human rater performance.
• Automated volumetry of SAH pathologies from admission CT scans can lead to novel imaging biomarkers.
• This approach holds promise for improved outcome prediction in patients with subarachnoid hemorrhage.