Impact of Intraventricular Hemorrhage on Classication of Hematoma Expansion and Development of Radiomics Prediction Models

Background: To investigate the impact of intraventricular hemorrhage (IVH) on the classication of hematoma expansion (HE), and the development of radiomics models using features extracted from the baseline hematoma to predict HE. Methods: Eighty-four patients with baseline and follow-up non-contrast CT within 4~24 hours were included. The intraparenchymal hemorrhage (IPH) and IVH were separately outlined by an experienced neuroradiologist. HE was dened as an absolute hematoma growth >6 mL or percentage growth >33%. HE was determined based on two criteria, using IPH alone (HE P ) or IPH+IVH (HE P+V ). The radiomics analysis was performed by using PyRadiomics to extract features, followed by random forest algorithm to select features, and lastly the decision tree to build classication models. Results: The classication of expansion showed 37 (44%) HE P and 47 (56%) non-HE P based on IPH alone, and similar results of 38 (45%) HE P+V and 46 (55%) non-HE P+V based on IPH+IVH. The majority, >94% of HE patients, had a poor outcome (death or mRS>3 at discharge). Three radiomics analysis (RA) models were built. The rst model using baseline IPH to predict HE P (RA P-P ) showed an accuracy of 80% but loss of correlation with the clinical outcome; the second model using IPH+IVH to predict HE P (RA PV-V ) had a slightly higher accuracy of 81% and resumed the poor outcome association with HE; and the third model using IPH+IVH to predict HE P+V (RA PV-PV ) had the highest accuracy of 86% with preserved clinical outcome correlation of HE. The sensitivity, specicity, and accuracy of three decision trees (RA P-P , RA PV-P , RA PV-PV ) were 0.8/ 0.68/ 0.89; 0.81/ 0.92/ 0.72 and 0.86/ 0.82/ 0.89, respectively. Conclusions: The proposed radiomics approach with additional IVH information could be used to classify the expansion status highly associated with the clinical outcome and provide a robust tool for the enrollment of high-risk ICH cases in the anti-expansion trials.


Introduction
Spontaneous intracerebral hemorrhage (sICH) accounts for about 7-15% of all strokes and carries a mortality rate of about 40%, with half of the fatalities occurring within the rst two days after an ictus (1)(2)(3). The high rate of early neurological deterioration after sICH is in part related to active bleeding that may proceed for hours after symptom onset (4). Among patients undergoing head CT scans within 3 hours of sICH onset, 28-38% have hematoma expansion (HE) on follow-up CT scans, with volume greater than one third compared with the hematoma volume on original CT scans (3,4). Furthermore, HE had been proved to be an independent predictor of clinical deterioration and poor outcome (3,(5)(6)(7).
Several radiological predictors on the baseline non-contrast CT (NCCT) for HE had been proposed, such as hematoma volume, shape, hypodensities, density heterogeneity…etc(8-13). The pattern of heterogeneity can be analyzed using the texture features extracted by the radiomics approach, which has been shown capable of capturing various agnostic features to aid-in HE prediction (14)(15)(16)(17)(18)(19). The radiomics features could be further combined with clinical (19) and radiological variables(16, 17) to improve HE prediction accuracy.
Except for hematoma expansion in the brain parenchyma, the presence of intraventricular hemorrhage (IVH) at the baseline CT scan has been shown associated with mortality in patients with sICH (3,20,21). It was quoted as one risk factor in the ICH score (20), a clinical grading scale for risk strati cation of sICH.
Studies have reported that 30 to 50% of sICH patients experienced additional IVH (21). Recently, similar to HE, IVH expansion at follow-up CT has also been identi ed as a strong predictor of poor clinical outcome (22). It was shown that the inclusion of IVH expansion into the de nition of HE can improve the overall prediction accuracy of the 90-day outcome (23). Nevertheless, the IVH information was usually ignored in the conventional radiomics models using texture analysis (14)(15)(16)(17)(18)(19).
In this retrospective study, we aimed to investigate the impact of IVH on the radiomic analysis for HE prediction as compared to the conventional approach based on intraparenchymal hemorrhage (IPH).
Three different radiomics analyses were performed: 1) using IPH to predict expansion de ned based on IPH, noted as RA p−p ; 2) using IPH+IVH to predict the traditional expansion de ned based on IPH, noted as RA PV−P ; 3) using IPH+IVH to predict expansion de ned based on IPH+IVH, noted as RA PV−PV . The prediction performance was compared.

Study population
This study was approved by the Institutional Review Board of our hospital. The requirement to obtain informed consent was waived due to its retrospective nature. We reviewed the sICH database from the picture archiving and communication system (PACS) to identify the patients who underwent a baseline and a F/U NCCT within 4-24 hours from January 2014 to June 2018. In total, 119 patients were identi ed. The exclusion criteria included: 1) the co-existence of vascular lesions and brain tumor diagnosed during the same admission (N=10); 2) pediatric patients <18 years old (N=2); 3) patients who underwent brain surgery before follow-up CT (N=19); 4) patients with primary IVH and minimal IPH (N=4). Thus, a total of 84 patients (61males, 23 females; mean age 60.1± 12.4 years; range 34-94 years) were included in the analysis. Clinical information, including blood pressure (SBP>180 or <180mmHg; DBP>100 or <100mmHg) (24), bleeding diathesis (INR > 1.5, aPTT ratio >1.5 or platelet count < 1x10 5 /ml)(25), Glasgow Coma Scale (GCS)(20) at admission (13~15 or < 13), and mRS(26) at discharge (≤3 or >3) were collected. The in-hospital mortality and modi ed Rankin Scale (mRS) at discharge, were used as the outcome.
CT Imaging Protocol The brain CT was acquired using our standard protocol on a 64-slice CT (De nition AS; Siemens Medical Solutions, Forchheim, Germany). The scanning range was from the skull base to the cranial vertex with the following parameters: 120 kVp, 380 mAs, and slice thickness/spacing of 4.8/4.8 mm.

Manual Hematoma Segmentation and HE de nition
The segmentation of the ICH region of interest (ROI) was performed manually, using Image J (National Institutes of Health, Bethesda, MD). The ROI drawing for baseline and F/U CT of each patient was done in the same seating by a neuroradiologist(TCW with 14 years of experience). The IPH and IVH were outlined separately, to form two datasets: ICH P containing the ROIs of IPH; and ICH P+V containing the ROIs of IPH and IVH. Based on the hematoma volumetric change between baseline and F/U CT studies, HE was de ned as an absolute hematoma growth >6 mL or relative growth of >33% from the baseline ICH (5,27). For ICH P+V , there has no consensus de nition of expansion, so the same criteria were applied. After the HE status was de ned, the baseline ROIs of ICH P and ICH P+V were used to extract radiomics features, followed by feature selection and model building to predict HE.

Feature Extraction
The radiomics analysis (RA) procedures are illustrated in Figure 1.
For the ICH P or ICH P+V in one patient, all segmented ROIs on different slices were combined to form a 3D lesion mask, and the linear interpolation was utilized to convert the hematoma ROI to be isotropic. Then, a total of 107 features were calculated using the PyRadiomics, including 14 shape, 18 rst-order, 24 Gray Level Co-occurrence Matrix texture, 14 Gray Level Dependence Matrix texture, 16 Gray Level Run Length Matrix texture, 16 Gray Level Size Zone Matrix texture, and 5 Neighboring Gray Tone Difference Matrix texture. To select robust features, two separate lesion ROI drawing was performed in 30 randomly selected cases. The extracted features from two ROIs of the same lesion were correlated to calculate the intraclass correlation coe cient (ICC). Only features with ICC > 0.8 were considered in the subsequent analysis to build models.

Feature Selection and Decision Tree Model
The extracted features from baseline ICH P or ICH P+V were used to build radiomics models to predict expansion, using two de nitions for HE P and HE P+V . Three different analyses were performed: 1) using IPH to predict HE P , noted as RA P−P ; 2) using IPH+IVH to predict HE P , noted as RA PV−P ; 3) using IPH+IVH to predict HE P+V , noted as RA PV−PV . In this study, we applied the random forest algorithm to estimate the feature importance as the selection criteria, by permutation of out-of-bag feature observation. The bootstrap-aggregated decision trees(28) were used to evaluate the importance of these features in differentiating patients with and without HE. The signi cance of one selected feature could be assessed according to the decreased accuracy after this speci ed feature was removed. All features were sorted based on their importance, and then the different number of features starting from the top 1, 2, 3… was used to test their classi cation performance by using the binary decision tree. The split of the tree was based on the improvement of the cross-entropy. For each node, the cross-entropy of the classi cation results was calculated using the following formula: In which k is the number of classes and pi is the proportion of cases belonging to class i. For all of the parent and child nodes, the splitting of the nodes was determined by the threshold minimizing the crossentropy. The random forest and decision tree analysis were implemented using MATLAB 2019b.

Statistical Analysis
Statistical analyses of the clinical parameters were performed using the SPSS for Windows (V.24.0, IBM, Chicago, Illinois, USA). Discrete variables were presented as counts(n) and percentages(%), and continuous variables were presented as medians and interquartile ranges (IQR). Chi-square test and student t-test were performed for categorical and continuous data respectively. P values < 0.05 was considered statistically signi cant.

Hematoma Expansion Status De ned Using IPH (HE P )
The baseline ICH volume, change of ICH volume at F/U, and short-term outcome of all 84 patients are summarized in Table 1. The expansion results of IPH+IVH are also included in Table 1. When using the same criteria of total volume change of >6 mL or relative growth of >33% to de ne the expansion, 38 patients (45%) were HE P+V and 46 patients (55%) were non-HE P+V . When compared with the HE P classi cation result, there were three crossover cases. One patient with HE P was re-classi ed as non-HE P+V (Figure 2a), and two patients with non-HE P were re-classi ed as HE P+V (Figure 2b). All three survived the episode and were discharged from the hospital with mRS of 4 and 5. Although the status of three patients was changed, the difference between HE P+V and non-HE P+V remained the same as those reported between HE P and non-HE P , and expanders had higher in-hospital mortality (39.5% vs 6.5%, p<0.001), and overall poor outcome (94.7 vs. 69.6%, p=0.004).

HE Prediction Performance of Different Radiomics Models
Three radiomics models were built using the ICH ROI on the baseline to predict hematoma expansion. The results are summarized in Table 2. The decision tree of the third model built using the radiomics features of ICH P+V to predict HE P+V is illustrated in Figure 3.
In the rst model using the traditional IPH to predict HE P , i.e. RA P−P , there were 25 true positive (TP), 42 true negative (TN), 5 false positive (FP), and 12 false negative (FN) cases. The accuracy, sensitivity and speci city were 80%, 68%, and 89%, respectively. When replacing the hematoma ROIs of ICH P with ICH P+V for feature selection to predict HE P , i.e. RA PV−P , the accuracy was slightly improved to 81% with 34 TP, 34 TN, 13 FP, and 3 FN cases. The sensitivity of this prediction model was markedly increased to 92% with compromised speci city to 72%. When the hematoma ROIs of ICH P+V were used to predict HE P+V , i.e.
RA PV−PV , the prediction accuracy was further improved to 86% with 31 TP, 41 TN, 5 FP, and 7 FN cases.
The sensitivity and speci city were 82% and 89%, respectively. Figure 4a shows a case example, which is classi ed as an expander using both criteria. The model built using IPH alone (RA P−P ) gives false negative results, but the other two models based on IPH+IVH give true positive results and correctly predict this patient as an expander. Figure 4b shows another expander case, which is very rare that all three models fail and give false negative results. This patient had a very large hematoma and appeared to be homogeneous on CT, which might be the reason for the false prediction.

Radiologic parameters and early outcome of different decision trees
The comparison of the radiologic parameters and early outcome between the hematoma expanders and non-expanders of each decision tree was summarized in Table 3. As compared with the 48 non-expanders labelled by RA PV−PV , the 36 expanders had signi cantly shorter CT follow-up intervals, larger hematoma volume change, more brain surgery, higher in-hospital mortality and poor functional outcome at discharge. This nding was consistent with the results of original de nition of hematoma expansion (Table 1). On the other hand, there was no signi cant difference of CT follow-up interval and functional outcome at discharge between 30 expanders and 54 non-expanders classi ed by RA P−P . It could be attributed to the modest sensitivity (68%) of RA P−P for hematoma expansion. With hematoma ROIs of ICH P+V for feature selection, a total of 47 hematoma expanders tagged by RA PV−P resumed the correlation between hematoma expansion and poor outcome at discharge. However, the CT follow-up interval between the tagged expanders and non-expanders by RA PV−P showed no signi cant difference.

Discussions
In this study, we investigated the impact of IVH on the radiomic analysis for hematoma expansion by implementation of random forest algorithm using different HE de nition and hematoma ROIs of IPH without or with addition of IVH. We compared the prediction performance and clinical outcome correlation of three decision trees (RA P−P , RA PV−P and RA PV−PV ) in a small case series of 84 SICH patients. Using hematoma ROIs of both IVH and IPH for the revised HE de nition and decision tree builtup, RA PV−PV had best prediction performance with preserved clinical outcome correlation of HE. By addition of IVH for feature selection, RA P−PV could improve the sensitivity and resume the outcome correlation of HE prediction model (RA P−P ) using the traditional approach based on IPH per se.
Due to the relatively small sample size, we employed the random forest algorithm to extract imaging features and then implemented a binary decision tree to build the classi cation model. As a proof of principle study in a small dataset, this approach was more likely to yield satisfactory performance (34). In the conventional analysis using IPH to predict IPH expansion (RA P−P ), our result was comparable to the published studies (14,17,19). Using the combined baseline IPH+IVH to predict the revised HE P+V evaluated by combing both, i.e. the RA PV−PV , it showed the highest accuracy of 0.86, with the sensitivity of 0.82 and speci city of 0.89. The results suggest that the additional information of IVH could improve the performance of radiomics analysis for HE prediction. In most patients, the precise separation of IVH and IPH could only be performed manually with subjective judgment based on brain anatomy. The model developed using the combined IPH+IVH can be easily implemented by using an automatic computerbased segmentation tool (35)(36)(37). For patients with a high risk of expansion, more aggressive procedures, including immediate surgery, may be considered. Another very helpful clinical application of the HE prediction model is to identify eligible subjects who are likely to show HE to participate in anti-expansion drug trials for sICH (13,31). For this application, a high speci city is preferred. That is, patients who are unlikely to show HE should not be enrolled, to maximize the power of testing the drug e cacy by using the smallest number of subjects.
The initial presence of IVH was not associated with HE in our study. It was also found in the PREDICT study (25)and the cohort study of the BAT score(31) for ICH expansion prediction. However, IVH had been demonstrated as a risk factor of HE in a case series of 259 putaminal hemorrhage(38) and the INTERACT study (33). On the other hand, dynamic IVH change, including new IVH (15 cases) and any IVH growth (38 cases) were signi cantly associated with hematoma expansion in our study (P<0.001). This nding was consistent with previous studies (21)(22)(23). With respect to the early outcome, IVH at the baseline CT scan is also associated with mortality and poor functional outcome with crude OR of 4.2 and 4.1, respectively.
Considering the impact of IVH on the outcome prediction, and the relationship between the dynamic IVH change and hematoma expansion, addition of IVH ROI into the radiomics analysis might enrich the hematoma feature and strengthen the correlation between clinical outcome and the radiomic prediction model, that was re ected in the improvement of prediction performance of RA PV−P and RA PV−PV compared to RA P−P .
There are several limitations. First, this study was a retrospective design from a single-center with small sample size. Second, the request of the F/U CT scan was at the clinician's discretion, most likely due to the large baseline ICH and/or worsening symptoms. Consequently, there was a relatively high percentage of patients with hematoma expansion (37/84; 44%), and poor outcomes for almost all expanders (> 94%). Third, the radiomics models were built using the random forest algorithm and the decision tree method, without going through cross-validation. Therefore, this should be considered as a pilot study mainly for proof of principle, to demonstrate the feasibility of the analysis based on combined IPH+IVH. In the future, the AI software may be applied to automatically segment IPH and IVH on the baseline and F/U NCCT, to e ciently process a large number of patients and evaluate the clinical role of the developed radiomics prediction models.  Figure 1 The radiomics analysis owchart to build the ICH expansion model. The IPH and IVH are segmented by manual tracing of the hematoma on baseline and follow-up CT images. The absolute or percentage volumetric change is calculated to determine whether the patient is an expander, or a non-expander based on IPH or IPH+IVH. The baseline ROI is used to extract radiomics features, and then the important features are selected by using the random forest algorithm to build the prediction model with the decision tree.

Figure 2
Illustration of two crossover cases. (a) A 94-year-old male with right cerebellar hemorrhage is classi ed as an expander based on IPH (3.7 to 5.6 ml, 51% growth), but is reclassi ed as a non-expander based on IPH+IVH (5.8 to 7.6 ml, 31% growth < 33% threshold). This patient is discharged on Day-74 after ICH with mRS of 5. (b) A 52-year-old female with right thalamic hemorrhage is classi ed as a non-expander based on IPH (16.1 to 18.1 ml), but is re-classi ed as an expander based on IPH+IVH (23.7 to 31.0 ml, 7.3 ml growth > 6 ml threshold). This patient is discharged on Day-70 after ICH with mRS of 5.

Figure 3
The decision tree built for the RAPV-PV model, by using radiomics features extracted from IPH+IVH to predict expansion determined based on the IPH+IVH criteria. The nal results show 31 true positive, 41 true negative, 5 false positive, 7 false negative, with an overall accuracy of 86%.