Our interpretations are limited by the modest number of patients included, alongside the retrospective design of the study. Furthermore, the statistics presented here are exploratory, although the significance of the associations presented remained so after multiple comparison corrections. Additionally, we describe several steps to demonstrate that regression models fitted were robust. Although our fitted and internally validated regression model demonstrates good performance on our single-centre dataset, further work with larger, multi-centre datasets will be required to cross-validate and confirm the findings reported above. However, the results nonetheless highlight important and unexplored potential areas of further study within the vasospasm literature, and our presented analysis pipeline can easily be extended to larger datasets, prospective studies, collaborative segmentations, and more sophisticated statistical models.

The modified Fisher score was validated for prediction of delayed cerebral ischaemia[8] rather than radiological vasospasm, and so may not be expected to predict radiological vasospasm more accurately than blood volume. DCI remains challenging to diagnose, especially in sedated or high-grade patients[44], and unsurprisingly its onset is often missed[23]. However, early angiographic vasospasm is significantly associated with the subsequent development of DCI[45,46]. Further, vasospasm as detected on TCD or CTA is well correlated with clinical deficits[21,22,45,46], and both vasospasm duration and severity can be used to assess the likelihood of delayed cerebral ischaemia and therefore the need for clinical interventions such as angioplasty[21,24,47,48]. Therefore, we propose that radiological detection of vasospasm, as an objective biomarker, may provide important information that may guide clinical decision making. Total blood volume provides useful information that can be used to predict the likelihood of significant vasospasm, and identify a subpopulation of patients that may require more stringent monitoring and/or intervention. Further modelling would be necessary to determine the influence of CSF drainage factors which may assist in reducing subarachnoid blood volume.

Fully manual segmentations drawn out by experts or trained raters, despite being considered the gold standard method, are time- and labour-intensive, often require lengthy training periods[49], and risk introducing substantial intra-rater and inter-rater variability and bias arising from various sources, including differences in opinion about the ground truth segmentation and in anatomical knowledge about relevant structures[50]. Accordingly, previous work has shown substantial variation in consistency and reproducibility of manual segmentations across both raters and structures[51–53]. Over recent years, semi-automated and automated methods have begun to match and even outperform manual segmentation in metrics of precision and inter-rater variability across a variety of modalities and structures[49,52,54,55].

Our semi-automated method adds to the growing literature of potential applications for machine learning methods in radiological interpretation and triage, and removes some of this intra- and inter-rater variability. Many computer-assisted methods for delineation of blood volume have previously focused on segmentation of haemorrhages within non-subarachnoid space [39,56–59]. However, application of these methods to aSAH has been noted to be challenging [40,57], and accordingly Dice scores for segmentations of subarachnoid blood have been consistently lower than for other haemorrhage subtypes [37,53,60,61], and convolutional networks used to automatically segment intracranial haemorrhage that includes subarachnoid blood have only achieved low-to-moderate Dice scores [62,63].

We attempted to mitigate any bias on behalf of the rater through expert assessment and correction, and correspondingly our mean Dice score between original and corrected segmentation was high (0.92), indicating excellent agreement between rater segmentation and expert opinion that required minimal correction. This score was substantially larger than those comparing manual segmentations by different observers [53,61], and in previous literature [37,53,60,61]. In particular, Boers et al. [53] previously utilised similar methods to segment aSAH, but achieved only moderate Dice scores for its automated segmentations (mean Dice score 0.55, range 0.00 - 0.83). Nonetheless, some variability remains, as labelling training data for the classifier and placement of seeds remain as manual steps that may lead to unintentional biases in volumes.

Segmentations took on average 10-15 minutes per scan for the rater to perform. Although this is notably faster than manual segmentations, this remains a non-trivial period of time that may bottleneck scan reporting, and may therefore be unfeasible in clinical radiology contextS.

Potential extensions to a fully automated pipeline (such as nnU-Net[64], which has already seen use in brain tumour segmentation [65,66]) would address a number of these limitations and allow for development of a clinically valuable toolkit. Validated, fully automated segmentations would allow for faster, reproducible and accurate delineation of blood distribution and volume, and may remove variability introduced in manual steps. Nonetheless, these methods still require training and validation, and this dataset serves as an important repository in facilitating this research.