Recurrent Sub-volume Analysis of Head CT Scans for the Detection of Intracranial Hemorrhage

Abstract

Non-contrast head/brain CT (NCHCT) is the initial imaging study of choice in patients visiting any emergency services and could be the only investigation to guide management in patients with head trauma or stroke symptoms. Immediate preliminary radiology reports to trigger appropriate level of care is paramount in the emergency department. Our proposed solution comprises of an efficient method for the detection of intracranial hemorrhage, by creating multiple 2-dimensional (2D) composite images of sub-volumes from the original scan. We also propose a recurrent neural network which combines the sub-volume features and takes into consideration the contextual information across sub-volumes to give a scan level prediction. We achieve an overall AUROC of 0.914.