Deep-Learning-Based Thrombus Localization and Segmentation in Patients with Posterior Circulation Stroke.

Area of Science:

• Medical Imaging
• Neurology
• Artificial Intelligence

Background:

• Thrombus volume in posterior circulation stroke (PCS) impacts patient outcomes.
• Manual thrombus segmentation is time-consuming and impractical for large-scale studies.
• Automated methods are needed for efficient analysis of imaging characteristics in PCS.

Purpose of the Study:

• To develop the first automatic method for thrombus localization and segmentation on CT scans in patients with PCS.
• To evaluate the performance of a novel convolutional neural network (CNN) named Polar-UNet.
• To compare Polar-UNet with a baseline CNN (BL-UNet) and assess the impact of volume-based removal (VBR).

Main Methods:

• A multi-center retrospective study included 187 patients with PCS from the MR CLEAN Registry.
• A CNN (Polar-UNet) was developed to segment thrombi, restricting the volume-of-interest (VOI) to the brainstem.
• Volume-based removal (VBR) was implemented to reduce false positive localizations.
• Performance was evaluated using intra-class correlation coefficient (ICC), precision, recall, and Dice coefficient.

Main Results:

• Polar-UNet achieved a thrombus localization recall of 0.82 without VBR, outperforming BL-UNet (0.78).
• VBR significantly improved precision to 0.65 for Polar-UNet and 0.56 for BL-UNet, with a minor recall reduction.
• Polar-UNet achieved a higher Dice coefficient (0.44) compared to BL-UNet with VBR (0.38).
• Both methods showed an ICC of 0.41 for automated vs. manual thrombus volumes.

Conclusions:

• Restricting the VOI to the brainstem in Polar-UNet enhances thrombus localization precision, recall, and segmentation overlap.
• The developed automatic method provides a practical tool for large-scale analysis of thrombus characteristics in PCS.
• VBR effectively improves precision in thrombus localization for automated CT segmentation in PCS.