Improving Deep Learning Automated Intra-operative Assessment of Reperfusion During Mechanical Thrombectomy Using Complementary Quantitative Angiographic Maps. (16th November 2020)

• Record Type: Journal Article
• Authors: Bhurwani, Mohammad; Snyder, Kenneth V; Waqas, Muhammad; Mokin, Maxim; Rava, Ryan; Podgorsak, Alexander; Davies, Jason; Levy, Elad I; Siddiqui, Adnan H; Ionita, Ciprian
• Is Part Of: Neurosurgery. Volume 67(2010)Supplement 1
• Publication Date: 2020-11-16
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1093/neuros/nyaa447_226
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract: INTRODUCTION: Modified-TICI (mTICI) scores are commonly used to provide assessment of cerebral tissue reperfusion during mechanical thrombectomies of large vessel occlusion (LVOs) acute ischemic strokes. Currently, manual analysis of angiograms is required to provide an mTICI score which may lead to inconsistencies. To improve consistency and optimize clinical workflow, an intra-operative algorithm using deep learning (DL) and a single quantitative angiography map was developed to automatically assess level of tissue reperfusion. This algorithm performed with an accuracy of 78.2%. Quantitative angiography, however, provides multiple maps that may contain complimentary information regarding level of tissue reperfusion. METHODS: Three hundred eighty-three angiograms from patients undergoing thrombectomies of anterior circulation LVOs were collected and assigned mTICI scores. Two quantitative angiographic maps, peak height (PH) and area under the time density curve (AUC), were derived from the capillary angiographic phase. These maps were masked to exclude regions outside the cranial cavity. A DL algorithm was developed to classify angiograms into 2 outcomes, mTICI 0, 1, 2a or mTICI 2b, 2c, 3. Sub group analysis was conducted to assess network performance using PH or AUC maps independently, and using both maps with an ensemble network. DL performance was evaluated and compared using classification accuracy (CA), area under receiver operating characteristic curve (AUROC), Matthews correlation coefficient (MCC) and two-tailed McNemar's test p-values. RESULTS: Results for PH, AUC and both maps ensembled were respectively: Average CAs, 78.2% (95% CI: 76.9%-79.5%), 77.2% (76.1%-78.4%), and 79.6% (78.5%-80.7%); Average AUROCs, 0.85 (0.83-0.86), 0.84 (0.82-0.85), and 0.85 (0.83-0.86); Average MCCs, 0.56 (0.53-0.58), 0.54 (0.51-0.56), and 0.59 (0.56-0.61). McNemar's tests indicated no significant advantage to using either maps by themselves (p-value > 0.05), but a significant advantage to using both maps ensembled over using either map independently (p-values < 0.05). CONCLUSION: This study proves feasibility of using multiple complimentary quantitative angiographic maps to improve performance of DL algorithms to provide real-time assessment of level of reperfusion during mechanical thrombectomy procedures.