Automated LVO detection and collateral scoring on CTA using a 3D self-configuring object detection network: a...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) is increasingly used for automated medical image analysis.
• Accurate detection of large vessel occlusion (LVO) and collateral scoring on computed tomography angiography (CTA) are crucial for stroke management.

Purpose of the Study:

• To evaluate the performance of a novel DL model, nnDetection, for automated LVO detection and collateral scoring on CTA.
• To assess the model's accuracy and agreement with expert radiologists.

Main Methods:

• A multi-task 3D object detection network (nnDetection) was trained on 2425 single-phase CTA scans from five centers.
• The model's performance was validated on an external test set of 345 CTA scans.
• Ground-truth labels for LVO and collateral scores were established by three radiologists.

Main Results:

• The nnDetection model achieved 98.26% diagnostic accuracy in identifying LVO, correctly classifying 339 out of 345 scans.
• DL-based collateral scores demonstrated good agreement with radiologist consensus (kappa = 0.80).

Conclusions:

• The self-configuring 3D nnDetection model accurately detects LVO on single-phase CTA.
• This DL approach provides reliable collateral scoring, supporting automated stroke diagnostics in LVO patients.