Detection of Acute Intracranial Haemorrhage in CT Scans Using Deep Learning Algorithms

Intracranial hemorrhage (ICH) refers to bleeding within the skull, affecting millions of individuals. It encompasses various subtypes based on their brain location, including epidural hemorrhage (EDH), subdural hemorrhage (SDH), subarachnoid hemorrhage (SAH), intraventricular hemorrhage (IVH), and intraparenchymal hemorrhage (IPH). Prompt detection and management of ICH are crucial, as it is a life-threatening emergency with high morbidity and mortality rates. Computed Tomography (CT) imaging is the primary tool for identifying and diagnosing neurological disorders. Rapid and accurate diagnosis of the presence, type, and location of an ICH is essential for timely medical intervention. However, the complexity of brain anatomy and variability in hemorrhage appearance make this task challenging. MobileNetV3 Large is an ideal CNN model for mobile devices because of its reduced number of parameters and compact model size. However, it requires improvements to better capture critical features. On the other hand, Xception, derived from InceptionV3, is highly effective at feature extraction with fewer but highly efficient parameters. This research proposes a double-branch model based on the Xception and MobileNet architectures. This model extracts spatial and temporal features, concatenating them and feeding them into a decision tree classifier for final predictions. The project involves training and validating deep networks from both Xception and MobileNet V3 Large branches. Our results indicate that the double-branch Xception-MobileNet V3 Large model outperforms both the double-branch Xception and the double-branch MobileNet V3 Large models individually. The performance of the double-branch Xception and the double-branch Xception-MobileNet models is approximately the same, highlighting Xception’s significant contribution to overall performance. Meanwhile, MobileNet V3 Large contributes complementary features that enhance the model's generalizability. The combined model strikes a balance across various metrics, demonstrating superior performance.