Intracranial Hemorrhage (ICH) is an acute brain stroke condition associated with high mortality and low functional recovery rate. Computed Tomography (CT) scans are most widely used to definitively draw a diagnosis for ICH. However, depending on the location and extent of the neurological damage, patients condition can be classified based on the modified Rankin Scale (mRS) of daily life dependability. Patient health outcome is very crucial but very difficult to predict automatically from the brain CT scans. In our work, we study an ensemble architecture of three state-of-the-art convolutional neural networks (CNNs), namely the VGG16, InceptionV3, and ResNet50 models. We utilize these deep learning models to detect ICH in the patient brain CT scans, and will eventually also predict the overall patient health outcome based on the mRs. We obtained the highest accuracy of 94.26% on a recent publicly available dataset (CQ500) that consists of 491 3D CT scans with 193,317 slices. The proposed strategy of automatically identifying the cerebral hemorrhage and predicting the individual outcome could assist the radiologists and neurologists in automating the triage process and expediting the decision making about the individual patient's treatment method.
Slice selection from 3D brain CT scans in cases with and without ICH
B. Iyer, S. Deoghare, et al. Automatic intracranial hemorrhage detection from brain CT scans using ensemble CNNs. In preparation, 2022.
B. Iyer, S. Deoghare, S. Hacker, V. Khandwala, D. Wang, D. Woo, A. S. Vagal, V. B. S. Prasath. Predicting ICH patient outcome from brain CT scans using an ensemble deep learning framework. Advanced Computational Neuroscience Network (ACNN), University of Michigan, Ann Arbor, MI, USA, 19 - 20 September, 2019.