Validation of a deep-learning based thrombus classifier on digital subtraction angiography using a large-scale...

Area of Science:

• Neuroradiology
• Medical Imaging Analysis
• Artificial Intelligence in Medicine

Background:

• Digital subtraction angiography (DSA) interpretation is subjective and observer-dependent.
• Mechanical recanalization is a key treatment for anterior circulation occlusions.
• Clinical decision-support tools are needed to improve diagnostic accuracy in neuroradiology.

Purpose of the Study:

• To evaluate the diagnostic performance of a deep-learning (DL) based thrombus classifier.
• To assess the utility of the DL model as a clinical decision-support tool for mechanical recanalization.
• To determine the sensitivity and specificity of the DL classifier on a large dataset of DSA images.

Main Methods:

• Retrospective analysis of an in-house dataset of DSA image series from endovascular recanalization procedures.
• Inclusion of DSA runs before and after recanalization for 309 patients (1,236 series).
• Application of an AI system to classify thrombus presence and assessment of diagnostic performance metrics (sensitivity, specificity, false-positive rate).

Main Results:

• The DL classifier demonstrated high sensitivity (87.6%) for proximal vessel occlusions (M1/M2 segments).
• Sensitivity was substantially lower for distal occlusions (M3/M4 segments: 23.1%) and anterior cerebral artery occlusions (27.3%).
• Overall specificity for thrombus detection was 89.8% with 17 false-positive classifications.

Conclusions:

• The DL classifier shows high sensitivity for proximal vessel occlusions on DSA, confirming its potential utility.
• The classifier's low sensitivity for distal occlusions indicates a need for further model development and training.
• Future work will focus on improving the DL model's performance for detecting distal occlusions prior to clinical implementation.