Detecting hemorrhage types and bounding box of hemorrhage by deep learning

Intracranial hemorrhage (ICH) a major health problem and the most common imaging method in ICH is computed tomography (CT). Detecting and locating it can help clinicians with diagnosis. Although deep learning models are well suited for detection and segmentation applications, detecting a bounding box may be employed as a major step to increase the segmentation accuracy. Localizing hematoma via bounding box is easier than semantic segmentation that tries to segment pixel-wise. Moreover, it is thought that as the bounding box shows the location, height, and width of the hemorrhage, it may also provide enough information to the doctor for diagnosis or to help the experts to detect anomalies in the brain. To evaluate and validate the proposed approach, Brain Hemorrhage Extended (BHX) dataset was employed. Intraparenchymal, Subarachnoid, Intraventricular, Epidural, Subdural, Chronic Subdural Hematoma were detected and enclosed in bounding box by using recently published YOLOv4 deep learning model. 94%, 92%, and 93% were achieved for overall precision, recall, and F1 score respectively by the proposed approach.