Abstract WP92: Deep Learning Approach for Diagnosis of Hemorrhagic Transformation in MRI: A Pilot Study

Abstract

Introduction: Several successful mechanical thrombectomy (MT) clinical trials in acute ischemic stroke (AIS) patients have dramatically increased the use of MT, which is also associated with increased risk of Hemorrhagic Transformation (HT). Yet, the reliability and agreement across multiple raters using the ECASS-II HT classification has been mixed. In this study, we introduce a residual 3D CNN for classification of HT in MRI gradient recalled echo (GRE) images. Methods: Ninety-nine patients diagnosed with AIS, treated with MT, and imaged with a 3.0T MRI scanner 24 hours post MT were included. After pre-processing, GRE images were separated into training (79), validation (10), and test (10) sets, all consisting of comparable numbers of no HT versus HT. Presence and classification of HT for all 99 patients using ECASS-II criteria were documented initially by the treating clinical team and reviewed by an independent image reader. Results: The model achieved 100% accuracy on the training set, 80% on the validation set, and 70% on the test set. The algorithm reliably classified parenchymal HT but erred in some cases of petechial HT or identified similar imaging features such as large vessels or microbleeds (Figure). Conclusions: Developments on this algorithm could lead to an automated tool for real-time identification of HT using acute stroke MRI. Figure: Images and automated prediction for the test set: A) accurate prediction of HT presence, B) accurate prediction of HT absence, and C) inaccurate prediction in which the algorithm predicted No HT in a case with both acute and chronic HT on contralateral sides, No HT in a case with small HT, and HT in a case with a particularly prominent thrombus.