Identification of successful cerebral reperfusions (mTICI ≥2b) using an artificial intelligence strategy.

ABSTRACT

BACKGROUND: The modified thrombolysis in cerebral infarction (mTICI) scale is a widely used and validated qualitative tool to evaluate angiographic intracerebral inflow following endovascular thrombectomy (EVT). We validated a machine-learning (ML) algorithm to grade digital subtraction angiograms (DSA) using the mTICI scale. MATERIALS AND METHODS: We included angiograms of identified middle cerebral artery (MCA) occlusions who underwent EVT. The complete DSA sequences were preprocessed and normalized. We created three convolutional neural networks to classify DSA into two outcomes, low- (mTICI 0,1,2a) and high-grade (mTICI 2b,2c,3). RESULTS: We included a total of 234 angiograms in this study. The area under the receiver operating characteristic was 0.863 (95% CI 0.816-0.909), 0.914 (95% CI 0.876-0.951), and 0.890 (95% CI 0.848-0.932) for the anteroposterior (AP), lateral (L), and combined models, respectively, when dichotomizing outcomes into low and high grade. The models' area under the precision-recall curve was 0.879 (95% CI 0.829-0.930), 0.906 (95% CI 0.844-0.968), and 0.887 (95% CI 0.834-0.941) for the AP, L, and combined models. CONCLUSION: In complete cerebral DSA, our angiography-based ML strategy was able to predict mTICI scores following EVT rapidly and reliably for MCA occlusions.