Masked Multi-Task Network for Case-Level Intracranial Hemorrhage Classification in Brain CT Volumes

We propose a novel Masked Multi-Task Network (MMT-Net) to detect brain CT volumes with intracranial hemorrhage and distinguish hemorrhage type(s) using only case-level labels. Different types of intracranial hemorrhage (The five types of intracranial hemorrhage (ICH) are intraparenchymal hemorrhage (IPH), intraventricular hemorrhage (IVH), subarachnoid hemorrhage (SAH), subdural hemorrhage (SDH) and epidural hemorrhage (EDH).) are defined by their anatomical locations. To utilize the brain structural information that is relevant to types of intracranial hemorrhage, brain masks were extracted during image preprocessing using a pre-trained brain CT segmentation network. Regional brain masks were then constructed for the central (cBrain) and the peripheral (pBrain) parts of the brain. These masks were later used as the input and the ground-truth brain masks to train the neural network. We designed a new two-branch network that encoded region-related features. The features were then fed into multi-task classifiers, which predicted both the regional brain masks and the region-related hemorrhage types. We also used the message passing module based on the conditional random field (CRF) model to refine features. We trained and tested our MMT-Net with a large in-house clinical dataset, and demonstrated superior performance of MMT-Net compared with the baseline ResNeXt50 network with the squeeze-and-excitation module. When tested using the 2019 RSNA intracranial hemorrhage challenge dataset, our MMT-Net trained with case-level labels more accurately detected hemorrhage cases and classified hemorrhage types than the challenge winner.