Deep Learning Approach to Predict Intracerebral Hemorrhage and Grad-Cam Visualization on CT Images

Imaging technology has the potential to achieve positive outcomes in rapid diagnosis and treatment plans for acute neurological diseases such as intracerebral hemorrhage (ICH) which is a severe type of stroke. In recent days, convolutional neural networks (CNN) based deep learning techniques have been recognized as a powerful tool for image recognition in detecting ICH lesions. In this study, we have proposed a CNN model with convolutional and fully connected layers to extract the informative features and classify the lesion automatically using 3293 computed tomography (CT) images of the brain. We achieved improved results with accuracy, recall, precision, and F1-score of 99.39%, 99.69%, 99.09%, and 99.4% respectively for classifying ICH lesions. In this study, Grad-CAM) was employed to spot regions of input CT images that lead most significantly to model predictions. These results indicate that the deep learning (DL) approach has the potential for accurate detection of hemorrhage stroke with better accuracy at a faster rate and thus can assist radiologists in clinical diagnosis.