Automated segmentation of subarachnoid hemorrhages with convolutional neural networks

Purpose: To investigate the viability of convolutional neural networks (CNNs) for the detection and volumetric segmentation of subarachnoid hemorrhage (SAH) in non-contrast computed tomography (NCCT). Materials and methods: We developed and trained a CNN for the SAH segmentation by splitting a set of 302 baseline NCCTs into a training (268) and a validation set (34). Segmentation accuracy was assessed on an additional 473 baseline NCCTs of SAH patients by calculating the intraclass correlation coefficient of the SAH volume and the Dice coefficient of the segmentations. We subsequently evaluated whether the developed SAH segmentation network can be used to discriminate SAH from acute ischemic stroke using 280 scans to optimize the discrimination and 70 scans for testing. Additionally, we tested whether the CNN-based volumetric SAH segmentation can also be used for hemorrhage segmentation in 396 NCCTs of rebleed patients. Results: The SAH volume agreement was high with an intraclass correlation coefficient of 0.966. The average Dice coefficient of the volumetric SAH segmentation was 0.63 ± 0.16, which is similar to expert interobserver agreement. The differentiation of SAH from ischemic stroke patients achieved an accuracy of 0.96. Despite the common presence of severe metal artifacts in scans of rebleed patients due to coiling, the CNN-based segmentation appears to be suitable for segmentation of rebleeds as well with comparable accuracy. The average CNN detection and segmentation processing time was 30 s. Conclusion: The proposed CNN is fast and accurate in detecting and segmenting SAH in NCCT scans.