Objective: Hemorrhagic transformation (HT) is a major complication of reperfusion therapy in acute stroke. We aim to explore the feasibility of HT location prediction from baseline magnetic resonance perfusion (MRP) source images.
Methods: Consecutive acute ischemic stroke patients who had HT after reperfusion therapy (IV tPA and/or endovascular thrombectomy) were reviewed from two stroke centers. Patients with MRP and diffusion-weighted imaging (DWI) before and 24h after treatment and follow-up susceptibility weighted imaging (SWI) were included. The HT was depicted on follow-up SWI semi-automatically. The DWI lesions and HT were classified into 12 cerebral vascular territories (Fig 1A). Then a total of 80,000 tissue-voxels were extracted from both HT region and non-HT region randomly with the ratio of 1:1. Each voxel included a set of values of MRP source image and baseline DWI. Convolutional neural networks (CNN) were trained with 3-fold cross-validation. Prediction maps were generated from the CNN for each patient. The prevalence of HT after infarction and the sensitivity and specificity of HT prediction in each vascular territory were calculated.
Results: Seventy-seven HT patients were analyzed (40 male, age 73 ±14 years, median baseline NIHSS score 14 [IQR 8-19], median onset-to-treatment time 242 min [IQR 152-355], median modified Rankin score 3 [IQR 1-4], 28 with Parenchymal hematoma). HT occurred most frequently in the territory of middle cerebral artery superficial branches (67.5%) and basal ganglia (37.6%), and least frequently in brain stem or thalamus (3.9%). HT was not evenly distributed throughout the vascular territories (Fig 1A). The best model was a locally-connected CNN (area under curve of 0.88). The average sensitivity and specificity for HT location prediction in anterior circulation were 89% and 60% (Fig 1BC).
Conclusion: The CNN trained from DWI and MRP source images could predict the HT locations in acute ischemic stroke patients.