Deep Learning based Automated Screening for Intracranial Hemorrhages and GRAD-CAM Visualizations on Non-Contrast Head Computed Tomography Volumes

Abstract

Intracranial Hemorrhage is a serious medical emer-gency which requires immediate medical attention. With most of the countries facing acute shortage of radiologists, it is important to develop an automated system which analyses the radiographic images and prioritise cases that require urgent medical attention. In this context, there has been attempts to apply deep learning (DL) techniques to the Head Computed Tomography (CT) slices to detect hemorrhage adequately in the past, where annotation effort is spent for individual slices of the CT volume for building a model. Our work aims to develop a robust model for the annotated CT volume dataset, which does not require slice level information for the presence of hemorrhage so that the annotation effort could be cut down substantially. A novel DL pipeline architecture based on the combination of convolutional neural network (CNN) and bi-directional long-short-term-memory (biLSTM) to capture both intra and inter slice level features for diagnosing hemorrhage from the non-contrast head CT volumes is introduced. The proposed model achieved a high accuracy score of 98.15 %, specificity of 1, sensitivity of 0.96 and F1 score of 0.98 with 95.3 % mitigation in the labelling effort of radiologists. However the performance scores are very well comparable to the scores achieved by the state-of-the-art models trained over the CT Volumes with slice wise annotation pertaining to intracranial hemorrhage detection. Additionally, the novel contribution is in integrating Gradient-weighted Class Activation Mapping (GRAD-CAM) visualization to the system, to offer visual explanations for the decisions made and provide supplementary information forming a strong advocate to radiologists in the clinical evaluation stage. The novel system is a first step towards building a robust autonomous assistive technology for radiologists, and leads to develop similar pipelined DL architecture for extracting information pertaining to other neurological disorders from Non-Contrast Head CT volumes.