Detection of Intracranial Hemorrhage Using Deep Learning

Abstract

Intracranial hemorrhage has always been a crucial medical condition where bleeding within the cranium region occurs, leading to severe neurological damage and sudden demise of a person. A patient's likelihood of survival in the treatment of Intracranial Hemorrhage is dependent on rapid diagnosis based on the radiologist's assessment of Computed Tomography (CT) scans. As a result, the requirement for precise and prompt identification of Intracranial Hemorrhage (ICH) is of utmost importance. Deep learning models can be used to assist this process by accelerating the time it takes to identify them. We built a deep learning model which will accelerate the time required to identify intracranial hemorrhages such that it facilitates the classification and segmentation of Intracranial Hemorrhage. We have constructed an EfficientNetB4 model using a Convolutional Neural Network (CNN) architecture which was used for the classification of images. This model comprises various layers categorized into different types, including convolution layers, normalization layers, fully connected layers, and activation layers. And we used Grad-Cam model for the image segmentation process which generated heat map for the image that contained the hemorrhage. We have used approximately 700,000 DICOM files collected from four international universities by the Radiological Society of North America (RSNA). We achieved an accuracy of 97 percent with learning rate of 0. 000125, batch size of 32 and 15 epochs of model training.