Background
Patients with acute ischaemic stroke and occlusion of the proximal anterior cerebral circulation confirmed by CT/MR angiography, with no contraindications, are eligible for thrombectomy with or without thrombolysis.
Brainomix e-CTA is an artificial intelligence (AI) clinical decision support software designed to detect large vessel occlusion (LVO).
Objective
Our primary objective was to determine the overall accuracy of Brainomix e-CTA software in the identification of LVOs.
Our secondary objective was to compare the accuracy and reliability of Brainomix e-CTA in identifying internal carotid artery (ICA) and proximal middle cerebral artery (MCA) occlusions vs. distal MCA occlusions.
Methods
A single-centre retrospective analysis of 410 consecutive patient records with CT angiography imaging was undertaken, spanning 15 months. 110 cases were excluded due to no Brainomix, non-diagnostic CTA study or no CTA performed. 300 cases were included for analysis. Intra- and extracranial vessels were reviewed independently on each CT angiogram study and comparison was made between reported findings and Brainomix findings. True positive, false positive and false negative cases were selected for further analysis.
The efficacy of Brainomix e-CTA in identification of ICA/MCA occlusion was calculated in one general category and two further subcategories: 1. Overall score 2. ICA/proximal MCA and 3. Distal MCA.
Precision and recall scores were calculated and used to derive the F1 measure (weighted average) for each category.
Results
1. Overall score: precision score of 55.2% and recall score of 77.1%. The F1 measure was 64.3%.
2. ICA/proximal MCA: precision score of 68.9% and recall score of 86.1%. The F1 measure was 76.5%.
3. Distal MCA: precision score of 27.2% and recall score of 50%. The F1 measure was 35.2%.
Brainomix e-CTA overcalled in approximately 45% of all cases and missed just under 1 in 4 positive cases.
It missed 14% of ICA/proximal MCA LVOs and overcalled 31% of cases.
Precision and recall were much lower in distal MCA occlusions, where it overcalled in roughly 73% of cases and missed 50% of positive cases.
Conclusions
Overall, Brainomix e-CTA software is a helpful adjunct in the identification of LVO in acute ischaemic stroke, its overall accuracy being 64%. It is more reliable and sensitive in identifying ICA/proximal MCA LVO as compared with distal MCA occlusions (accuracy scores of 76.5% and 35.2% respectively).