Deep Learning-based Multiclass Segmentation in Aneurysmal Subarachnoid Hemorrhage

IntroductionAneurysmal subarachnoid hemorrhage (aSAH) is a life-threatening condition with a significant variability in patients outcomes. Radiographic scores used to assess the extent of SAH or other potentially outcome-relevant pathologies are limited by interrater variability and do not utilize all available information from the imaging. Image segmentation plays an important role in extracting relevant information from images by enabling precise identification and delineation of objects or regions of interest. Thus, segmentation offers the potential for automatization of score assessments and downstream outcome prediction using precise volumetric information. Our study aims to develop a deep learning model that enables automated multiclass segmentation of structures and pathologies relevant for aSAH outcome prediction. MethodsOut of 408 patients treated with aSAH in the department of Neurosurgery at Charite University Hospital in Berlin from 2009 to 2015, a subset of 73 representative CT scans were included in our retrospective study. All non-contrast CT scans (NCCT) were manually segmented to create a ground truth. For the multiclass segmentation task we determined six different target classes: basal and cortical SAH, intraventricular hemorrhage (IVH), ventricles, intracerebral hemorrhage (ICH), aneurysms and subdural hematoma (SDH). An additional hemorrhage class was created by merging the individual hemorrhage classes. The set of 73 NCCT was splitted into three stratified sets: training set (43 patients), validation set (10 patients) and test set (20 patients). We used the nnUnet deep learning based biomedical image segmentation tool and implemented 2d and 3d configurations. Additionally, we performed an interrater reliability analysis for multiclass segmentation and assessed the generalizability of the model on an external dataset of primary ICH patients (n=104). Segmentation performance was evaluated using: median Dice coefficient, volumetric similarity and sensitivity. Additionally, a global Dice coefficient was calculated by considering all patients in the test set to be one single concatenated image. ResultsThe nnUnet-based segmentation model demonstrated performance closely matching the interrater reliability observed between two senior human raters for the SAH, ventricles, ICH classes and overall hemorrhage segmentation. For the hemorrhage segmentation a global Dice coefficient of 0.730 was achieved by the 3d model and a global Dice coefficient of 0.736 was achieved by the 2d model. The global Dice coefficient of the SAH class was 0.686 for both of the nnUnet models; ICH: 0.743 (3d model), 0.765 (2d model); ventricles: 0.875 (3d model), 0.872 (2d model). In the IVH, aneurysm and SDH classes the nnUnet models performance differed the most from the human level performance. Overall, the interrater reliability of the SAH class was observed to be lower than in other classes. In the external test set a global Dice coefficient of 0.838 for the hemorrhage segmentation was achieved. ConclusionDeep learning enables automated multiclass segmentation of aSAH-related pathologies and achieves performance approaching that of a human rater. This enables automatized volumetries of pathologies identified on admission CTs in aSAH patients potentially leading to imaging biomarkers for improved aSAH outcome prediction.