An Interpretable Deep Learning System for Automatic Intracranial Hemorrhage Diagnosis with CT Image

Abstract

Intracranial Hemorrhage (ICH), a dangerous and devastating medical emergency, affects thousands of patients every year around the world. In the clinical settings, Computer Tomography (CT), is widely used for diagnosis of neurological diseases. In the situation of Intracranial Hemorrhage, not only saving time is critically important, but also the expertise to accurately diagnose and locate ICH is imperative. However, there are not always enough doctors working in the emergency expert in the field of ICH, and the results from using only deep learning models are not always reliable. Three neural networks, VGG-19, Resnet-101, and DenseNet-201 were trained separately on preprocessed the Intracranial hemorrhage data with labels and used the Grad-CAM method to produce a saliency map by visualizing the process of the network making a decision regarding to specific class index, thus increasing the interpretability of the results. We tested the networks' performances on our preprocessed CT data, and their differences produced saliency maps. Three experiments were designed and conducted to help us understand our models' performance and predictions in different contexts. First, we observed the differences between the pre-trained deep learning model and the unpre-trained deep learning models. Second, we observed how the performance and Grad-CAM results would differ when the images were normalized at different Window values. Third, we merged the six Grad-CAM images generated by the six class indices for each image into a single image and fed it into the network to observe the results. To further demonstrate the potential application of our deep learning models, we used trained models to make a GUI software called ICH Deep Learning Detector in python with the PyQt5 library to simplify the process of doctors using the deep learning model and learning from predictions.