Applying Quantitative Radiographic Image Markers to Predict Clinical Complications After Aneurysmal Subarachnoid Hemorrhage: A Pilot Study

Abstract

Accurately predicting clinical outcome of aneurysmal subarachnoid hemorrhage (aSAH) patients is difficult. The purpose of this study was to develop and test a new fully-automated computer-aided detection (CAD) scheme of brain computed tomography (CT) images to predict prognosis of aSAH patients. A retrospective dataset of 59 aSAH patients was assembled. Each patient had 2 sets of CT images acquired at admission and prior-to-discharge. CAD scheme was applied to segment intracranial brain regions into four subregions, namely, cerebrospinal fluid (CSF), white matter (WM), gray matter (GM), and leaked extraparenchymal blood (EPB), respectively. CAD then detects sulci and computes 9 image features related to 5 volumes of the segmented sulci, EPB, CSF, WM, and GM and 4 volumetrical ratios to sulci. Subsequently, applying a leave-one-case-out cross-validation method embedded with a principal component analysis (PCA) algorithm to generate optimal feature vector, 16 support vector machine (SVM) models were built using CT images acquired either at admission or prior-to-discharge to predict each of eight clinically relevant parameters commonly used to assess patients’ prognosis. Finally, a receiver operating characteristics (ROC) method was used to evaluate SVM model performance. Areas under ROC curves of 16 SVM models range from 0.62±0.07 to 0.86±0.07. In general, SVM models trained using CT images acquired at admission yielded higher accuracy to predict short-term clinical outcomes, while SVM models trained using CT images acquired prior-to-discharge demonstrated higher accuracy in predicting long-term clinical outcomes. This study demonstrates feasibility to predict prognosis of aSAH patients using new quantitative image markers generated by SVM models.

I. INTRODUCTION

Aneurysmal subarachnoid hemorrhage (aSAH) is a medical emergency associated with high mortality, morbidity, and significant healthcare burden⁶. Nearly 30,000 people are affected by aSAH every year in the United States¹¹. Recent studies suggest that early brain injury (EBI) and delayed cerebral ischemia (DCI) resulting from several pathophysiological processes are major determinants of mortality and morbidity associated with aSAH patients. EBI typically occurs within 72 hours of aSAH, while DCI is commonly seen 4–14 days of aSAH onset. Thus, early and quick diagnosis and prediction of aSAH prognosis is important to more effectively treat aSAH patients in clinical practice.

EBI occurs secondary to one or more of the following: acute hydrocephalus (HCP), transient cerebral oligemia, and systemic inflammatory response. EBI causes significant morbidity and mortality¹³. Despite this known associative role of EBI in determining the clinical outcome, the absence of objective radiological image biomarkers limits its role in assessing disease severity and predicting prognosis. The commonly used modified Fisher scale (mFS) identifies the amount of blood in subarachnoid space and ventricles. Although easy to use and able to predict cerebral vasospasm and DCI, it is subjective with high inter-rater variability^14,8. Quantifying blood using the above-mentioned parameters has a decent ability to predict the occurrence of DCI and long-term clinical outcome, but it is semi-automated and also operator dependent¹⁸. In addition, recent studies reported DCI as a complex process, not fully explained by intracranial blood or HCP¹⁶. Hence, researchers have investigated if EBI could be radiologically assessed through global cerebral edema (GCE), which might incorporate some of the above-mentioned complex processes to predict clinical outcomes more objectively. Recently, a semi-quantitative non-automated scoring system has been investigated to assess GCE using Subarachnoid Hemorrhage Early Brain Edema Score (SEBES) based on sulci’s visibility at two pre-determined levels, which may predict DCI with high accuracy². Additionally, quantification of cerebral edema using semi-automated or automated segmentation of sulci volumes has shown to be an important marker for analyzing EBI after aSAH^5,22.

As described above, in order to detect or quantify the severity of aSAH and predict its prognosis or clinical outcomes, non-contrast computed tomography (CT) imaging is typically applied to scan patients’ brain due to the high efficiency of CT imaging scan to detect leaked blood on brain tissue. However, subjectively reading brain CT images by neuroradiologists is difficult with large inter-reader variability in assessing EBI and hence prone to predictive errors. It is likely that many useful imaging features depicting on brain CT images acquired at the admission of patients and subsequent CT scans during hospitalization may or may not be correctly identified through visual inspection and subjective interpretation of neuroradiologists alone. Applying an automated program to detect and segment brain CT images has the potential to provide a standardized and unbiased estimation of targeted radiological image parameters. Hence, in this exploratory study, we developed and tested a novel computer-aided detection (CAD) scheme or automated radiographic imaging software aiming to quantify image features that pathophysiologically contribute to short-term (in-hospital) and long-term clinical outcomes. Successful development of such CAD-generated radiological image biomarkers will help in the early prediction of possible clinical outcomes of aSAH patients.

The hypothesis used to develop current image analysis software is based on the two studies^2,5 that used an automated program to segment and quantify a pre-determined level above lateral ventricle to perform a volumetric assessment of sulci, white matter (WM), gray matter (GM), and extra-parenchymal blood (EPB). Hence, the objective of this study was to develop and test a new CAD scheme to automatically segment brain regions and generate radiographic imaging biomarkers computed from brain CT images to predict various clinical outcomes after aSAH.

II. MATERIALS AND METHODS

2.1. Study Population

This study was approved by the Institutional Review Board of the University of Oklahoma Health Sciences Center (OUHSC). All patients used as study subjects had previously signed written informed consent forms for their participation in the long-term follow-up portion of the study during their 1-year clinical follow-up, as discussed in our previous studies^17,18. In brief, CT scans of 59 patients admitted with the diagnosis of aSAH at the OUHSC between the years 2014 and 2016 were retrospectively analyzed for this study. The inclusion criteria for the study subjects required patients diagnosed with aSAH and a CT scan acquired within 48 hours of symptom onset. Multiple head CT scans were acquired to different clinical indications during the hospital course. However, we only used two CT image scans acquired during patients’ admission and day 10–14 after hospitalization for this study, which are termed as the first and last CT scans in the rest of this paper. The 10–14 days were chosen as it is generally accepted to be the end of the DCI period (prior-to-discharge).

Following 4 short-term clinical outcomes during hospitalization were retrospectively studied, which include: (i) Clinical vasospasm (CVSM) – defined as patients developing clinical symptoms attributable to cerebral vasospasm after exclusion of other confounding causes, (ii) DCI – defined as the occurrence of CVSM and/or radiological evidence of cerebral infarction, (iii) HCP – assessed on admission CT scan using age-corrected bicaudate index¹⁰, and (iv) ventriculoperitoneal shunt (VPS) – placement during or after discharge as clinically determined. Additionally, clinical outcomes assessed during two post-discharge clinic visits (i.e., 3 months and one year) include two indices, namely, (1) a modified Rankin score (mRS) that assesses physical disability ranging 0 to 6 with higher values indicating worse outcome and (2) a Montreal cognitive assessment (MoCA) index that assesses cognitive disability which is scored 0–30 with lower score indicate the presence of cognitive dysfunction. These generate 4 long-term clinical outcomes defined in this study.

Due to the small dataset of 59 patients, we divided each clinical outcome into two categories of good or poor outcome, as discussed below. The clinical measures DCI, CVSM, HCP, and VPS, were categorized into two classes of either presence (poor) or absence (good). The mRS is an ordinal disability score categorized into seven scales (from 0 = no symptoms to 6 = dead). We divided mRS into two categories: 0 to 2, independent functioning, and 3 to 6 as dependent outcome or death. A MoCA score of 26 or greater is considered normal. Thus, we divided MoCA into two categories a MoCA score of 26 or over (normal cognition) and a MoCA score of less than 26 (presence of cognitive disability). The mRS and MoCA were performed in patients at ~3 months (3M) and ~1 year (1FU) follow-up visits. The distribution of the patients for the above eight measures along the demographics was summarized in Table 1.

Table 1.

Demographics and Clinical Characteristics of patients (n = 59).

Clinical Characteristics		Number of cases (n)
		First scan	Last scan
	Sex	Male: 20; Female:39	-
Age: (Min 20; Max 83)	Mean ± SD	52.7 ± 12.9	-
	Median (IQR)	53 (44, 60)
	DCI (Yes/No)	59 (33/26)	57 (33/24)
	CVSM (Yes/No)	59 (23/36)	57 (23/34)
	HCP (Yes/No)	39 (15/24)	37 (15/22)
	VPS (Yes/No)	53 (13/40)	51 (13/38)
	mRS at 3M (class-1/class-2)	41 (27/14)	40 (26/14)
	MOCA at 3M (class-1/class-2)	41 (19/22)	40 (19/21)
	mRS at 1FU (class-1/class-2)	36 (22/14)	34 (20/14)
	MOCA at 1FU (class-1/class-2)	36 (16/20)	34 (14/20)

mRS: class-1: 0–2; class-2: 3–6; MOCA: class-1: ≥ 26; class-2 < 26; SD – Standard Deviation; IQR – Interquartile range, 3M – 3 months, 1FU – 1 year follow-up.

2.2. Flow diagram of the proposed CAD scheme

CAD scheme is divided into two phases, namely, the image processing phase and the data analysis phase. The image processing phase is common, while the data analyses phase changes depending on the clinical measure being evaluated. During the image processing phase, each three-dimensional (3D) input brain CT scan is processed to (i) automatically segment the volumetric region of the intracranial brain, (ii) label each voxel to a different subcategory, and (iii) compute corresponding radiographic image features. While in the data analysis phase, the computed image features associated with a specific type of scan (first or last) coming from the image processing phase are combined with class labels of each clinical measure to perform classification analysis. In this study, we assessed the classification performance of our CAD scheme for each of eight clinical measures as shown in Table 1 independently using both first and last CT scans. As a result, 16 different classification analyses were performed belonging to 8 clinical measures for both the first and last CT scans. Figure 1 illustrates our CAD scheme’s graphic user interface (GUI) and the detailed flow diagram of the proposed CAD scheme with the highlighted individual steps in each phase. More detailed explanations of each step of the CAD scheme are discussed in the subsequent sections.

Figure 1.

Illustration of the graphic user interface (GUI) of our CAD scheme (a) and a detailed flow diagram of each step of the proposed CAD scheme categorized by the image processing and the data analysis phases (b).

2.3. 3D Brain Segmentation and CT Labeling

A two-stage consecutive mapping-based algorithm is implemented in this CAD scheme to achieve optimal segmentation results⁷. During the first stage, for each 3D CT scan sequence, the CAD scheme applies an initial image pre-processing step to identify three image markers, namely, global minimum (gm), local minimum (lm), and maximum peak (mp) of the intracranial brain region. These image markers will represent slice indices of the current CT sequence as reference points to guide the CAD segmentation algorithm. To achieve this, CAD performs Otsu thresholding to filter the largest single-connected intracranial brain region belonging to each of the 2D-image slices. Next, based on a comparison of segmented brain volumes, the slice with the largest volume is considered mp, the slice with the smallest volume is regarded as gm, and the slice with the smallest volume in the opposite direction concerning the mp slice gm slice is marked as lm. At the end of the first stage, the initial segmentation results are still prone to errors regarding leakages and identification of multiple brain regions per slice.

Thus, in the second stage, a fine-tuning technique in reference to the neighboring slices is applied to obtain precise segmentation. Correction in this stage begins with the mp slice as the intracranial brain region in this slice is the most significant, single connected component and wholly enclosed within the skull, facilitating accurate segmentation. Then, the fine-tuning procedure scans either in the left or right direction, applying the consecutive mapping technique. During this mapping-based correction, the previous slice’s segmentation results are used to define the limiting boundary criterion to avoid segmentation leakages and identify multiple connected brain regions. The limiting boundary criterion performs morphological dilation on prior segmentation results to determine current slice boundary limits, preventing segmentation leakages. In the case of multiple connected regions, each individual region is examined with prior segmentation results and constrained to enclose within the limiting boundary for inclusion in the current slice segmentation. To summarize the second stage of the segmentation method, the process begins and considers the segmentation at mp slice as the reference, then continues using this consecutive mapping-based algorithm until either a gm slice or lm slice is reached in both directions covering all the image slices.

After successful brain region segmentation, based on the standardized calibration of Hounsfield Units (HU) of CT images, each CT voxel inside the brain is subcategorized into four regions, namely, Cerebrospinal fluid (CSF), White Matter (WM), Grey Matter (GM), and blood using multi-thresholds of HU. In this study, the default threshold (TH) values defined for subcategorization of voxels were TH1: 20, TH2: 32, and TH3: 50. During the CT labeling protocol, any voxel with an HU value less than or equal to TH1 is identified as CSF. In contrast, all the voxels with HU value greater than TH3 are labeled as blood. Then, the voxels with HU values greater than TH1 and less than or equal to TH2 are labeled as WM. Finally, the voxels with HU values greater than TH2 and less than or equal to TH3 are labeled as GM. To summarize, CT labeling protocol: (1) CSF ≤ TH1, (2) TH1 < WM ≤ TH2, (3) TH2 < GM ≤ TH3, (4) TH3 < blood was applied to categorize each voxel into one of the four regions. Following CT Labelling, volumetric analysis of these four regions was performed to identify and compute radiographic image features. The analysis was conducted only on image slices between the level of lateral ventricles until the top of the skull. This pre-determined level was picked from the previous study², where the presence of sulci is analyzed using a qualitative assessment for association with aSAH. All the CSF regions automatically identified by the CAD scheme are considered as sulcal regions as CSF is filled within the sulcal space.

2.4. Computing radiographic image features

Next, CAD scheme is applied to compute the image features from CT images slice-by-slice. The three-dimensional (3D) volumetric representation of each image feature is then computed by weighted summation of feature values calculated in all corresponding CT image slices. For each CT label (i.e., Sulci, WM, GM, Blood, and Brain), an initial slice-based representation of volumes is computed using the number of the segmented voxels (N_Label) and slices parameters such as voxel length (V_L), voxel breadth (V_B), and slice thickness (S_T) as shown in equation 1.

$$Labe{l}_V{\text{|}}_{slice}=N_{LABEL}\times V_L\times V_B\times S_T{|}_{slice}$$ (1)

Next, these slice volumes were summated to yield 3D volumes of each labeled category, as shown in equation 2.

$$Labe{l}_V=\sum _{slic{e}_{k=m}}^{n}Labe{l}_V{\text{|}}_{slic{e}_k}$$ (2)

The above process is repeated independently for each CT label to obtain corresponding volumetric features. For instance, to compute the 3D volume of the labeled sulci, the CAD scheme first computes the segmented sulcal volumes depicting each slice separately. It then combines the results of all the slices to obtain the 3D volume of sulci, as shown in equations 3 and 4.

$$Sulc{i}_V{\text{|}}_{slice}=N_{Sulci}\times V_L\times V_B\times {S_T|}_{slice}$$ (3)

$$Sulc{i}_V=\sum _{slic{e}_{k=m}}^{n}Sulc{i}_V{\text{|}}_{slice}$$ (4)

In this way, the first group of features includes volumes of 5 labeled categories, namely, sulci, white matter (WM), gray matter (GM), blood, and brain tissue, are computed, which are represented using a “V” as subscript following each label as follows: (i) SULCI_V, (ii) WM_V, (iii) GM_V, (iv) BLOOD_V, and (v) BRAIN_V.

Additionally, the CAD scheme also computes the second group of features that includes 4 features obtained by calculating the proportion SULCI_V with respect to the volumes of remaining labels (WM_V, GM_V, BLOOD_V, and BRAIN_V) and represented with a “SULCI” as a subscript following each label as follows: (i) WM_SULCI, (ii) GM_SULCI, (iii) BLOOD_SULCI, and (iv) BRAIN_SULCI, respectively. For instance, the proportion of sulci with respect to WM is represented using WM_SULCI and is computed as shown in equation 5.

$$W{M}_{SULCI}=\frac{Sulc{i}_V}{W{M}_V}$$ (5)

As a result, a total of nine radiographic image features, including five features from the first group and four features from the second group, are extracted and computed from 3D CT slices of each case.

2.5. Building Machine Learning Model and Evaluation

We then build and test several multi-feature fusion-based machine learning (ML) models to classify the prognosis of the study cases into two classes. Specifically, we investigated eight clinical measures (as shown in Table 1) for two CT scans (first and last), resulting in a total of 16 different independent analyses of case classification. Although many different types of ML models can be used for this prediction purpose, we chose support vector machine (SVM) for the classification tasks in this study due to its simplicity and high robustness when applied to a relatively small image dataset. Based on our study experience¹², a polynomial kernel is used in all SVM models built in this study.

In order to build robust and highly performed ML models, the following three critical issues or challenges need to be considered and addressed, which are: (i) imbalance of cases between two classes, (ii) a scope for redundancy in the initial feature pool, and (iii) a relatively small dataset. First, there is a noticeable imbalance between the number of samples belonging to each class regarding each clinical measure in our dataset. For example, in the case of CVSM, we have a total of 59 samples with the first CT scan, and the number of samples belonging to class-1 vs. class-2 is 23(39%) vs. 36(61%) respectively. More details related to the class imbalance situation regarding each of the 16 analyses can be observed in Table 1. To address this issue and minimize its impact on the training ML model, we applied a Synthetic Minority Oversampling Technique (SMOTE) algorithm²¹ to rebalance the number of samples in each class. The vital point of applying SMOTE is that it introduces synthetic data samples by interpolation between some minority class instances within a specified neighborhood. The fundamental purpose of the SMOTE is to generate synthetic data samples belonging to the minority class label using the interpolation technique of minority class instances within a specified neighborhood. For instance, if we need a total of N new synthetic samples to balance the minority class, according to a distance metric, we first select K samples belonging to the minority class from the training set. Then, among these K instances, N instances are chosen randomly for computing the new instances by interpolation. A sample illustration of SMOTE can be seen in figure 2, in which u represents the minority class label, and [u1, u2, u3, u4] represents the N instances selected. In contrast, the points [v1, v2, v3, v4] represent the synthetic data generated using interpolation. Second, to potentially avoid overtraining and reduce the redundancy of the input features, we applied principal component analysis (PCA). The PCA is set to generate a new principal component feature vector with a variance rate of 95%. Third, a leave-one-case-out (LOCO) based cross-validation method is adopted to train and evaluate each ML model due to the small dataset. A LOCO validation method maximizes learning power while minimizing the case partition and testing bias, as demonstrated in previous studies (i.e.,^1,15).

Figure 2.

A sample illustration of generating synthetic data of the minority class label using the interpolation of SMOTE algorithm.

Then, we embedded both SMOTE and PCA algorithms into the LOCO cross-validation process during training and testing each SVM model. Specifically, in each LOCO training and testing iteration, we first take out one case from the training set as an independent testing case. Second, SMOTE algorithm is applied to generate synthetic data to balance the minimum class instances. Third, PCA is applied to the training samples with a variance rate of 95%, thereby generating a new feature vector of reduced length. Fourth, the SVM model is trained using all training samples (including the synthetic samples). Last, after the SVM model is trained, the model is applied to the testing case to generate a classification score. This SMOTE and PCA embedded LOCO training and testing iteration process is repeated 59 times for testing all cases in our dataset. As a result, 59 classification scores generated by each SVM model are recorded.

To evaluate the performance of each SVM model, we used the following two steps. First, a receiver operating characteristic curve (ROC) is constructed from the classification scores using a maximum likelihood-based ROC curve fitting program (ROCKIT, http://www-radiology.uchicago.edu/krl/, University of Chicago), and the area under the ROC curve (AUC) along with the standard deviation (STD) is computed and used as an index to evaluate and compare the performance of each ML model to classify between two classes. Second, we apply an operating threshold on the classification scores (T = 0.5) to divide all testing cases into two classes (score ≤ 0.5: ‘Class-1’; score > 0.5: ‘Class-2’). From the classification results, several confusion matrices corresponding to different SVM models are generated, which are used to compute various performance indices of class classification (i.e., classification accuracy (ACC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV)).

III. RESULTS

Figure 3 shows one sample of brain CT image segmentation results by applying the proposed consecutive mapping-based algorithm used in our CAD scheme. The sequence of images is organized from the top left to the bottom right in a left to right fashion. CAD scheme identifies 28 CT image slices depicting visible brain regions with varying sizes in the first 25 CT slices. In this case, CAD identifies slice-11 as the mp slice with the most significant brain area. In contrast, slice-1 and slice-25 are assigned as gm and lm, respectively. During the fine-tuning stage from mp towards the gm (left direction), we can notice that the guided limited boundary criteria defined from the prior slice segmentation have assisted in both avoiding segmentation leakages and identifying multiple connected brain regions (noticed in slices 1 through 7). Additionally, during the fine-tuning stage from mp towards the lm (right direction), when the top of the skull is reached at slice-26, segmentation is not performed for slices 26 through 28 as it is not required.

Figure 3.

A sample case of segmentation results using the proposed consecutive mapping-based algorithm. The sequence of images is organized from the top left to the bottom right in a left to right fashion.

As explained earlier, in this study, image data analysis is conducted only on image slices between the level of lateral ventricles until the top of the skull. Thus, the last slice index depicting the ventricles is marked for reference for each case. Then, all the images above the defined ventricle level until the top of the skull are selected for further analysis by CAD scheme. Figure 4 shows a sample series of images representing the CT labeled images within the defined region of analysis, including the reference slice consisting of ventricles. In Figure 4, slice-1 is only shown to illustrate the level of ventricles but is not included in further analysis. Here, the images are labeled in 4 greyscale or levels ranging from dark to bright, where each gray level represents CSF, WM, GM, and Blood, respectively.

Figure 4.

A sample case to illustrate CT labeled images within the defined region of analysis. The first slice in the sequence is for ventricles reference purpose only and not included in the volumetric analysis of labeled regions. In the figure, greyscale ranging from dark to bright in four levels represents cerebrospinal fluid (1), white matter (2), gray matter (3), and blood (4), respectively.

Figures 5 and 6 plot and compare the ROC curves of applying the SVM models to classify cases into two classes based on all eight clinical measures between the first and the last CT scan. The AUC values computed from the first CT scan range between 0.65±0.10 and 0.82±0.05, with the lowest value observed for classifying mRS at 1FU and the highest value for classifying DCI. Similarly, the AUC values for the last CT scan range between 0.62±0.07 and 0.86±0.07, with the lowest value observed for classifying DCI and the highest value observed for classifying MOCA at 1FU.

Figure 5.

A comparison of ROC curves for the short-term clinical measures between using the first and the last CT scan. Where (a) represents the curves plotted from the first CT scan, (b) represents the curves plotted from the last CT scan.

Figure 6.

A comparison of ROC curves for the long-term clinical measures between the first and the last CT scan. Where (a) represents the curves plotted from the first CT scan, (b) represents the curves plotted from the last CT scan.

For simplicity, we separated the clinical measures into two groups representing short-term and long-term clinical outcomes. The short-term clinical measures are the complications that occur during hospitalization, including DCI, CVSM, HCP, and VPS. Whereas the long-term clinical measures are observed at three months and 1-year follow-up after hospital discharge, including mRS at 3M, MOCA at 3M, mRS at 1FU, and MOCA at 1FU. In terms of short-term measures, on comparison between the first and the last CT scan, we observe that majority of these complications (including DCI, CVSM, HCP) are more accurately predicted using the first CT scan. In contrast, only VPS can be observed to have a higher accuracy using the last CT scan. On the contrary, to predict long-term measures (mRS and MOCA at 3M and 1FU), classification accuracy is higher using the last CT scan than using the first CT scan, which indicates the impact of treatment during hospitalization of the patients. More details regarding the best model selected for each of 16 data classification analyses and the performance metrics (AUC and ACC) are summarized in Table 2. A side-by-side comparison of AUC and ACC is demonstrated in Figure 7.

Table 2.

A comparison of classification performance metrics and the best model selected for each of eight clinical measures between the first and the last CT scan.

Scan Type	Clinical Measure	AUC±STD	ACC
First	DCI	0.82±0.05	0.78
	CVSM	0.76±0.06	0.78
	HCP	0.73±0.08	0.72
	VPS	0.77±0.07	0.77
	mRS at 3M	0.75±0.08	0.73
	MOCA at 3M	0.68±0.08	0.73
	mRS at 1FU	0.65±0.10	0.72
	MOCA at 1FU	0.73±0.08	0.64
Last	DCI	0.62±0.07	0.63
	CVSM	0.71±0.07	0.63
	HCP	0.68±0.09	0.68
	VPS	0.80±0.06	0.73
	mRS at 3M	0.85±0.07	0.80
	MOCA at 3M	0.79±0.07	0.73
	mRS at 1FU	0.71±0.09	0.62
	MOCA at 1FU	0.86±0.07	0.79

Figure 7.

A side-by-side comparison of both the performance evaluation metrics of AUC and classification accuracy for each of the 16 analyses. The ranges of STD are marked in AUC bars, and the detailed STD values are reported in Table 2.

From accuracy values (or bars as shown in Figure 7), we can notice that all the short-term measures are better performed using the first CT scan than the last CT scan. Whereas in the long-term measures, we observe that the majority of clinical measures or outcomes (including mRS at 3M, MOCA at 3M, and MOCA at 1FU) are better predicted using the last CT scan, and only mRS at 1FU is better predicted using the first CT scan. Last, based on 16 confusion matrices generated from the classification scores of their respective SVM models, a set of the additional computed performance indices, including sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), are summarized and compared as shown in Table 3.

Table 3.

A comparison of several other classification performance indices for each of eight clinical measures between the first and the last scan.

Scan Type	Clinical Measure	Sensitivity	Specificity	PPV	NPV
First	DCI	0.88	0.65	0.76	0.81
	CVSM	0.87	0.72	0.67	0.9
	HCP	0.73	0.71	0.61	0.81
	VPS	0.85	0.75	0.52	0.94
	mRS at 3M	0.96	0.29	0.72	0.80
	MOCA at 3M	0.63	0.82	0.75	0.72
	mRS at 1FU	0.91	0.43	0.71	0.75
	MOCA at 1FU	0.75	0.55	0.57	0.73
Last	DCI	0.79	0.42	0.65	0.59
	CVSM	0.83	0.5	0.53	0.81
	HCP	0.67	0.68	0.59	0.75
	VPS	0.77	0.71	0.48	0.9
	mRS at 3M	0.88	0.64	0.82	0.75
	MOCA at 3M	0.79	0.67	0.68	0.78
	mRS at 1FU	0.80	0.36	0.64	0.56
	MOCA at 1FU	0.86	0.75	0.71	0.88

Additionally, Table 4 shows two sets of AUC values along with the standard deviations of applying each of 9 single image features computed by our CAD scheme from the images of the first, and last CT scans to predict or classify mRS at 3-months among 41 aSAH patients. AUC values range from 0.52±0.10 to 0.73±0.09. The results indicate that many of these image features contain relatively higher levels of discriminatory power. For example, 4 image features yield AUC > 0.6 to predict mRS at 3-months using both CT images acquired at both the first and the last CT scans. However, the results also show that using an SVM model to optimally fuse multiple image features yields the highest classification performance (i.e., AUC = 0.85±0.07 using the images acquired from the last CT scan).

Table 4:

AUC values and standard deviations of applying each of 9 single image features computed by our CAD scheme to predict or classify mRS at 3-months using the images of the first and last CT scans.

Feature Number	First Scan	Last Scan
F1	0.66±0.10	0.57±0.09
F2	0.71±0.09	0.65±0.10
F3	0.72±0.09	0.73±0.08
F4	0.57±0.09	0.54±0.11
F5	0.64±0.09	0.58±0.10
F6	0.67±0.09	0.52±0.10
F7	0.56±0.10	0.72±0.08
F8	0.62±0.10	0.63±0.09
F9	0.62±0.10	0.66±0.09
Proposed Combined Model	0.75±0.08	0.85±0.07

IV. DISCUSSION

This study demonstrates that sequential brain CT scans can play an important objective role in quantitatively predicting both short-term and long-term clinical outcomes of aSAH patients. Although many clinical measures (i.e., 8 measures presented in Table 1) have been proposed and investigated, these measures are subjectively and qualitatively assessed by neuroradiologists in current clinical practice³, which is a difficult task with potentially higher inter-rater variability. However, we propose developing this novel unbiased quantitative image analysis software, which can provide objective and quantitative radiological image marker or data analysis tool to predict aSAH outcomes. In addition, since previous studies have demonstrated the advantages of using new image markers to facilitate clinical research and assist clinicians in their decision making process^4,9, we in this study investigate and demonstrate the feasibility of developing and applying a new CAD scheme that uses quantitative radiographic image biomarkers generated by multi-feature fusion-based machine learning (ML) models to predict the prognosis of various clinical measures commonly used in current clinical practice after aSAH. The quantitative image markers enable us to assess both the short-term and long-term clinical measures, which were predominantly observed between the first and last CT scans. To the best of our knowledge, no similar CAD scheme or ML-generated quantitative image markers have been developed and applied to predict a variety of short-term and long-term clinical outcome measures of aSAH patients.

This study has several unique characteristics and contributions to develop and apply novel quantitative image markers. First, the image processing pipeline of the CAD scheme is fully automated and capable of performing various tasks, including (i) accurate segmentation of intracranial brain volume using a novel mapping-based segmentation scheme, (ii) automated labeling to subcategorize each CT voxel into four clinically relevant brain regions, and (iii) computing relevant radiographic image features representing the volumetric characteristics of these defined brain sub-regions. Since all CAD steps mentioned above are performed automatically without user intervention, the procedure is less time-consuming and has higher reproducible or robust results. Additionally, the CAD scheme is also visually interactive along with various functionalities that allow users to either monitor or adjust volumetric estimations. Even though none of these features were used in this study, clinical professionals will examine them in future experiments to improve the segmentation markings. These corrections will provide a ground truth of CT labeling, thereby resulting in building more robust ML prediction or classification models.

Second, our study demonstrates the feasibility of identifying and applying radiographic image features or markers computed from brain CT images to potentially phenotype aSAH patients and predict the prognosis of various associated clinical measures. This study supports the radiomics concept and provides a new opportunity to expand the radiomics concept²⁰ to more broad clinical application fields. The study is influenced by an article² using qualitative markers evaluated by the clinicians as assessment in SAH patients and developed automatic quantitative image markers that are consistent and not prone to human error. Additionally, based on the results of this study, we have two observations, namely, (i) the image markers extracted and computed from the admission (first) CT scan have the potential to quickly predict the short-term clinical measures or outcome, which may help clinicians quickly decide optimal treatment options to improve clinical outcome, (ii) whereas the image markers extracted from the 10–14-day (last) CT scan can more accurately predict the long-term clinical measures in aSAH patients. Such observation clearly supports the impact of patients’ treatment on the long-term clinical outcome or recovery, which can assist clinicians in developing optimal rehabilitation plans for the patients discharged from the hospital. As a result, our study results and observation indicate the potential advantages of developing new quantitative image markers, which has promising potential to provide clinicians (neuroradiologists) new decision-making supporting tools to assist them optimally treating initial brain injury and also mitigating secondary brain injury to reduce the morbidity and mortality of aSAH patients.

Third, unlike many previous studies to identify image markers based on a single optimal image feature, we build and apply SVM models to generate multi-feature fusion-based quantitative image markers. The SVM models are built and integrated with an optimal multi-feature fusion method that enables to capture the complementary image phenotype information. As a result, using the SVM model yields a higher classification performance than using a single optimal feature, as shown in Table 4, which coincides with the previous studies^1,21. Additionally, many precautionary measures are taken while developing and training these ML models, namely, (i) SMOTE algorithm is used to address and solve the class imbalance issue, (ii) PCA method is applied to generate optimal and non-redundant feature vector to achieve feature dimensionality reduction with a variance rate of 95%, and (iii) LOCO cross-validation method is used to maximize learning power of available dataset and eliminate bias in case partition. Both SMOTE and PCA are also embedded into LOCO cross-validation to further reduce ML model training bias.

Fourth, although most previous studies used p-values or odds ratios to evaluate the association between image or clinical markers and patients’ prognosis or response to treatment, assessment results of association cannot be directly applied in the clinical practice to develop or establish precision medicine or personalized treatment paradigm. To overcome this limitation and increase clinical relevance, we use the ROC-based data analysis method and build confusion matrices to evaluate the performance or accuracy of the ML models to predict patients’ prognoses or clinical measures. The reported AUC values, along with classification sensitivity, specificity, and positive or negative predictive values (PPV or NPV), have higher clinical significance to assist clinicians (neuroradiologists) in making the optimal diagnosis and treatment decision applying to the individual patients in the clinical practice. Our study results with the highest AUC = 0.86±0.07 clearly demonstrate the feasibility or potential of developing and applying new quantitative image markers in future clinical practice.

Despite the encouraging results, we also recognize that this is only a pilot study with several limitations. First, the dataset assembled in this study is relatively small and may not sufficiently represent the varied spectrum of the aSAH patient population. Thus, a more extensive and diversified study cohort must be used to train/test a more robust CAD scheme. Second, simple thresholding is employed to perform CT labeling of various regions, which may not be ideal due to CT image noise. Although a more adaptive case-based thresholding may be investigated for better CT pixel labeling, as the increase of dataset, an advanced deep learning model based region segmentation method can also be developed and tested to segment brain CT images as demonstrated in one of our recent studies, which applies a modified deep leaning model to segment acute ischemic stroke (AIS) lesions using a publically available AIS dataset¹⁹. Third, we only compute and build a small feature pool with 9 features. As a dataset size increases, more radiomics types of image features or deep learning model generated automated features can be computed and investigated. Fourth, even though the interactive graphical user interface has multiple features to monitor/update segmentation and CT labeling, these steps were not considered and evaluated in this study. Thus, to overcome these limitations and improve ML model or image marker accuracy, in our future studies, we will ask experienced clinicians to validate/correct the possible errors in automated markings to improve region segmentation results, thereby improving the performance of prediction models.

In conclusion, despite several limitations, this is a valid pilot study that has clearly demonstrated the feasibility and potential advantages of developing new quantitative image markers to predict the prognosis of aSAH patients corresponding to various clinical measures used in current clinical practice. The initial results indicate the significance of both the admission (first) and day 10–14 (last) CT scans in their predictive capability of assessing both short-term and long-term clinical measures. Based on the foundation built in this study, more research efforts can follow to further optimize and validate the novel quantitative image markers by conducting prospectively clinical studies in the future.