Semantic Segmentation and Quantification of Subarachnoid Hemorrhage of Cerebral <scp>CT</scp> Using Deep Convolutional Neural Networks

ABSTRACT Fast and accurate evaluation of the extensiveness of the subarachnoid hemorrhage from computed tomography (CT) scans might facilitate timely and more accurate actions in the treatment of the condition. An automatic tool with the ability to integrate into the chain of actions of the radiologist would at least partially solve this issue. In this paper, we investigate a data set comprising patients ( N = 36) diagnosed with Subarachnoid Hemorrhage, collected from two distinct university hospitals. The contribution of this study lies in the modification of conventional segmentation models. Our approach involves augmenting the standard segmentation network architecture by introducing an additional regression head on top of the network. The findings reveal a considerable improvement in the overall performance of the proposed model compared to the standard use of deep learning for medical image segmentation: Intersection over Union (IoU) scores are 0.58 with an unmodified architecture versus 0.63 with our proposed enhanced network structure (and 0.64 vs. 0.70 for Dice coefficient). This augmentation not only enhances the detection accuracy of distinguishing SAH and non‐SAH layers but also results in more precise prediction of segmentation masks. The simultaneous consideration of both segmentation and regression tasks within the network leads to enhanced capabilities in localizing and quantifying SAH more accurately. Additionally, it is shown that the prediction of outcome (Glasgow Outcome Score) based on estimated SAH volume is at the same accuracy as if the hand‐segmented volumes are used: AUC was 0.73 with our segmentation results, while expert evaluation based AUC is 0.72.