Abstract No. 720 Identification of Irreversibly Damaged Brain Tissue on Computed Tomography Perfusion Using Convolutional Neural Network to Assist Selection for Mechanical Thrombectomy in Ischemic Stroke Patients
R. Wang,K. Chang,H. Zhou,J. Wu,G. Cohan,M. Jayaraman,R. Huang,J. Boxerman,L. Yang,F. Hui,J. Woo,H. Bai
DOI: https://doi.org/10.1016/j.jvir.2019.12.779
IF: 3.682
2020-01-01
Journal of Vascular and Interventional Radiology
Abstract:Endovascular treatment of ischemic stroke has shown positive clinical outcomes. Further optimization requires identifying patients who will benefit from reperfusion. We propose using deep learning, specifically 3D convolutional neural networks (CNN), to identify infarcted tissue (core) on CT perfusion (CTP) with diffusion weighted imaging (DWI) MRI as gold standard for irreversible brain infarction and evaluate lesion size impact on the network’s performance. Two sets of CTP and DWI maps: (1) 116 lesions from patients from a large academic center (LAC) who presented with ischemic stroke and (2) 94 lesions from the 2018 Ischemic Stroke Lesion Segmentation (ISLES) challenge were combined for the project. CTP maps (MTT, CBV, CBF, Tmax) were produced with RAPID software. The LAC cohort DWI were manually segmented using 3D Slicer by a neurologist. Preprocessing involved affine registration of the DWI to a CTP map using 3D Slicer followed by resampling to 1-mm isotropic. The CTP maps were normalized by dividing the intensity of each pixel by the max value across all patients. The data was split into training, validation, and test sets in a 7:2:1 split. A two-path CNN architecture from DeepMedic was trained with four input channels (MTT, CBV, CBF, Tmax) and binary output (infarct and non-infarct). Training was performed for 12 epochs with a batch size of 10, a learning rate of 0.0001, RMSProp optimizer, Nesterov momentum, and segment size of 37x37x37. The ISLES patients presented within 8 hours of stroke onset and underwent a DWI within 3hrs after CTP while LAC patients averaged 15 hours. Training was done on the combined dataset based on lesion diameter measured using all 210 lesions (94 ISLES, 116 LAC mean 57.5 vs. 32.3 mm, P <0.001), 138 lesions over 25 mm (86 ISLES, 52 LAC 61.0 vs. 51.3 mm, P = 0.059), and 89 lesions over 45 mm (66 ISLES, 23 LAC 75.7 vs. 74.2 mm, P = 0.833). The DICE scores of the lesions in the test set was 0.232 ± 0.053, 0.362 ± 0.0795, and 0.392 ± 0.077, respectively. Our preliminary results demonstrate that lesion size plays a key role in the performance of 3D CNN trained using CTP inputs to predict infarct core on DWI.