Intracranial Hemorrhage Detection in CT Scans using Deep Learning

In intracranial hemorrhage treatment patient mortality depends on prompt diagnosis based on a radiologist's assessment of CT scans. In this paper, we investigate the intracranial hemorrhage detection problem and built a deep learning model to accelerate the time used to identify the hemorrhages. To assist with this process, a deep learning model can be used to accelerate the time it takes to identify them. In particular, we built a convolutional neural network based on ResNet for the classification. Using 752,803 DICOM files collected from four international universities by the Radiological Society of North America (RSNA) [1], we trained and tested a ResNet-50 based model for predicting the hemorrhage type. Our model has an accuracy of 93.3% in making the correct multiclass prediction and an average per-class recall score of 76%. We show it is possible to achieve an average recall of 86% while maintaining 70% precision via tuning the prediction thresholds. Lastly, we show real-world applicability by deploying a simple web application. The source code for training, metrics evaluation and web application is available at [2].