Wessel Van Der Steen

Quantification of Subarachnoid Hemorrhage: Prediction of Delayed Cerebral Ischemia and Outcome

Aneurysmal subarachnoid hemorrhage (aSAH) is a severe form of stroke that occurs at a relatively young age and has a poor outcome compared with other forms of stroke. Delayed cerebral ischemia (DCI) is a complication of aSAH, which is associated with high mortality and a poor functional outcome. Early and accurate prediction of patients who will develop DCI or have a poor outcome could assist physicians in making treatment decisions and counseling patients and their families. Imaging characteristics, more specifically the amount of blood on CT-scan, have shown to be associated with the occurrence of DCI and poor clinical outcome. The aim of this thesis was to evaluate currently broadly utilized grading scales to estimate the amount of blood on CT-scan in patients with aSAH, introduce a new quantitative aSAH segmentation method based on convolutional neural networks and to develop new prediction models for DCI and clinical outcome using quantified blood volumes.

In Chapter 2, we performed a systematic review and meta-analysis on the association of currently broadly utilized radiological grading scales to assess the amount of blood after aSAH with clinical DCI. The assessed scales were the Fisher scale, modified Fisher scale, and Hijdra sum score. When possible, pooled odds ratios for the occurrence of DCI were calculated per grade increase on the radiological scale. Due to a large variability in data reporting and the varying definitions of DCI, only for the Fisher scale a meta-analysis could be performed. This analysis showed that of all 4 Fisher grades, Fisher grade 3 had the strongest association with DCI. However, in literature the modified Fisher scale was more commonly significantly associated with DCI than the Fisher scale. This advocates using the modified Fisher scale in favor of the two other reviewed scales.

In Chapter 3, we proposed a new automatic method for aSAH detection and volumetric segmentation based on convolutional neural networks (CNNs). A CNN was trained on 302 baseline non-contrast CT-scans and validated on 473 non-contrast CT-scans. The agreement between the automatically segmented blood volumes and manually delineated blood volumes was high (intraclass correlation coefficient of 0.966, Dice coefficient 0.63 ± 0.16). This accuracy is similar to expert interobserver agreement. Since the time for detecting and segmenting SAH was only 30 seconds on average, combined with its high accuracy, we believe that this method could be valuable in a clinical setting.

In Chapter 4, we investigated the association between the cisternal, intraventricular, intraparenchymal, and subdural blood volumes and the occurrence of DCI in patients with aSAH. Only the blood volume inside of the cisterns was significantly associated with DCI (adjusted OR 1.02 per milliliter blood volume). For the intraventricular, intraparenchymal, and subdural blood volumes no significant association was found. The findings in this chapter suggest that the amount of blood located inside of the cisterns plays a more important role in the development of DCI than the amount of blood in other locations of the brain.

In Chapter 5, we assessed the association of intraparenchymal hematoma volume and neurologic condition at admission with clinical outcome in patients with aSAH and a ruptured middle cerebral artery aneurysm. Furthermore, we evaluated the association of any treatment option (clipping, coiling, no treatment) with outcome. We found a significant difference in favorable outcome between patients with a poor (17% favorable outcome) and a good neurologic condition (68% favorable outcome) on admission. This significant difference was not found for large (> 50 ml) intraparenchymal hematoma volume (29% favorable outcome) and small (< 50 ml) hematoma volume (45% favorable outcome). Both in patients with small and large hematoma volumes no difference in outcome was found between patients that underwent clipping or coiling of the aneurysm, both with or without decompression and clot removal. Therefore, in this chapter we conclude that neurologic condition on admission plays a more important role in estimating the outcome of the patient than intraparenchymal hematoma volume in patients with ruptured middle cerebral artery aneurysms. Furthermore, as no difference in outcome was found between clipping and coiling of the aneurysm, this study suggests that the choice of coiling or clipping of the aneurysm in patients with a ruptured middle cerebral artery aneurysm with concurrent intraparenchymal hematoma is at the discretion of the local neurovascular treating team. In Chapter 6, we developed and internally validated a prediction model for clinical DCI in patients with aSAH including total blood volume (TBV) as one of the candidate predictors. From the prospective SAH registry of the Neurosurgery department of the Amsterdam University Medical Centers, 369 patients were included. Of all candidate predictors, only TBV was an independent predictor of DCI. The model including only TBV showed moderate predictive accuracy for DCI (c-statistic 0.64). Nevertheless, it performed better than a model including the Fisher (c-statistic 0.56) or modified Fisher scale (c-statistic 0.58). The findings in this chapter suggest that though blood volume plays a role in the development of DCI, other factors may need to be identified to achieve a higher accuracy for DCI prediction models. In Chapter 7, we aimed to improve DCI prediction by training a machine learning algorithm. A machine learning algorithm was trained using the same prospective dataset as in Chapter 6. Furthermore, automatically extracted image features were included in the model to account for blood volume and location. Combining the machine learning model trained on the prospective dataset and automatically extracted image features resulted in the highest predictive accuracy for DCI (c-statistic 0.74). Therefore, in this chapter we conclude that machine learning algorithms improve the prediction of DCI in patients with aSAH, particularly when image features are also included. In Chapter 8, we developed and externally validated a prediction model for clinical outcome in patients with aSAH including TBV as one of the candidate predictors. For developing the models, the same prospective dataset as in Chapter 6 was used. For external validation of the model, 317 patients were included from the prospective aSAH registry of the University Medical Center Utrecht. The TBV, neurologic condition, age, aneurysm size, and history of cardiovascular disease were independent predictors of outcome and were included in the final model. The externally validated predictive accuracy and discriminative power were high (R2 = 56% ± 1.8%; c-statistic = 0.89 ± 0.01). Including the cisternal-, intraventricular-, and intraparenchymal blood volume separately did not improve the model. Replacing the TBV with the modified Fisher scale significantly decreased the performance of the model. Therefore, in this chapter we conclude that the TBV-based prediction model for clinical outcome in patients with aSAH has a high predictive accuracy, significantly higher than a prediction model including the commonly used modified Fisher scale.