Deep Learning Assisted Image Data Analysis for Automatic Detection and Prediction of Cerebrovascular Diseases

Intracranial Artery Stenosis (ICAS), Cerebral Aneurysm, Acute Ischemic Stroke (AIS) and Acute Hemorrhagic Stroke (AHS) are very serious neurological diseases of brain that require emergency medical intervention. Doctors use varieties of images such as Computed Tomography (CT), Computed Tomographic Angiography (CTA), Magnetic Resonance Image (MRI), Magnetic Resonance Angiography (MRA), and Digital Subtraction Angiography (DSA) images to detect these brain diseases. Those collected images are highly unstructured and are stored in DICOM format. When the brain disease patients are admitted to a hospital, huge volumes of images are generated from the patients, which are usually unstructured. All these medical images of brain diseases are stored in the PACS of Chang Gung Memorial Hospital (CGMH) and Stroke Registry of Chang-Gung Healthcare System (SRICHS). To process and analyze these imaging data using classical statistics is very difficult and time-consuming as they comprise varieties of data types in different formats. Besides, these medical images are large in size and may contain 100 to 150 slices per patient. Analysis of various features in the images and correlating the features manually with the clinical data is a tedious job. Currently used classical statistical tools such as ANOVA, ANCOVA and MANCOVA are inefficient to handle the huge volume of the unstructured data for predicting the brain diseases. Hence, Deep Convolutional Neural Network (DCNN) is the most efficient and powerful technique by which different formats of images can be analyzed. The proposed THREE years project plan comprises the Brain stroke Neurologist and Neuro-Radiologist as the data team to support the domain knowledge of the brain diseases and respective clinical difficulties. Based on the advice of the doctors, we the Artificial Intelligence (AI) team would like to collect different formats of the retrospective image data through IRB license to analyze and design various prediction models using Deep Learning (DL). The objective of this proposal is to process and analyze the retrospective brain images using AI to design and establish the automatic outcome prediction models. In the 1st year of the plan, automatic classification and detection of the intracranial artery stenosis regions will be done by analyzing the MRA images of ICAS using DL. Besides, quantification of the stenosis region will be modeled in this year. In the 2nd year, retrospective DSA and NCCT image data of the Cerebral Aneurysm will be analyzed for classification of the intracranial hemorrhage (ICH) and subarachnoid hemorrhage (SAH) patients. Automatic prediction of SAH risk in un-ruptured cerebral aneurysm patients will be modeled using Deep Convolutional Neural Network. In the 3rd year, automatic localization models for ischemic and hemorrhagic stroke regions will be designed. Automatic prediction of ischemic stroke onset time and determination of hematoma progression models will be designed in this plan. All images and clinical data will be analyzed using Deep Convolutional Neural Network in TensorFlow, Keras and OpenCV frameworks and Alexnet, VGG 16, ResNet 50 and Inception V3 architecture. New trained models will be designed for the automatic localization and detection of the above mentioned neurological diseases by modifying the hidden layers.

Project ID:PB10907-2519
External Project ID:MOST109-2221-E182-014