Deep Learning based Automated Hemorrhage Segmentation and Volume Estimation from Computed Tomography Imaging

Intracranial hemorrhage, characterized by leakage of blood in the brain, has a high mortality and morbidity. Accurate volume quantification of blood in the brain is paramount for accurate risk assessment as well as treatment planning. In this study, we use a deep learning based approach to quantify the volume of blood in the brain from CT imaging. We evaluate a CT slice-based 2D segmentation model as well as an image volume-based 3D model. We quantified the accuracy of segmentation using Hausdorff distance and dice similarity coefficient (DSC). We quantified the overall performance as well as performance based on hemorrhage phenotype. We observed that the 3D volume-based model performed better than the 2D model (average DSC=0.79 vs 0.77). Subsequently, we also observed that the 3D model performed better in predominantly subdural hematoma (SH) cases (R²=0.94 vs 0.88) than in cases with intraventricular hemorrhage (IVH) (R²=0.87 vs 0.84). We found that although the 3D model had better segmentation performance, both models couldn't accurately capture blood volume in IVH cases.