Developing new quantitative CT image markers to predict prognosis of acute ischemic stroke patients

Abstract

BACKGROUND:

Endovascular mechanical thrombectomy (EMT) is an effective method to treat acute ischemic stroke (AIS) patients due to large vessel occlusion (LVO). However, stratifying AIS patients who can and cannot benefit from EMT remains a clinical challenge.

OBJECTIVE:

To develop a new quantitative image marker computed from pre-intervention computed tomography perfusion (CTP) images and evaluate its feasibility to predict clinical outcome among AIS patients undergoing EMT after diagnosis of LVO.

METHODS:

A retrospective dataset of 31 AIS patients with pre-intervention CTP images is assembled. A computer-aided detection (CAD) scheme is developed to pre-process CTP images of different scanning series for each study case, perform image segmentation, quantify contrast-enhanced blood volumes in bilateral cerebral hemispheres, and compute features related to asymmetrical cerebral blood flow patterns based on the cumulative cerebral blood flow curves of two hemispheres. Next, image markers based on a single optimal feature and machine learning (ML) models fused with multi-features are developed and tested to classify AIS cases into two classes of good and poor prognosis based on the Modified Rankin Scale. Performance of image markers is evaluated using the area under the ROC curve (AUC) and accuracy computed from the confusion matrix.

RESULTS:

The ML model using the neuroimaging features computed from the slopes of the subtracted cumulative blood flow curves between two cerebral hemispheres yields classification performance of AUC = 0.878±0.077 with an overall accuracy of 90.3%.

CONCLUSIONS:

This study demonstrates feasibility of developing a new quantitative imaging method and marker to predict AIS patients’ prognosis in the hyperacute stage, which can help clinicians optimally treat and manage AIS patients.

Keywords

Acute ischemic stroke (AIS)quantitative image markers prediction of AIS prognosis computer-aided detection and diagnosis (CAD)

1 Introduction

Stroke is the fifth leading cause of death, with more than 140,000 deaths each year in the USA [1]. Ischemic stroke is the most common stroke accounting for about 87% of all stroke cases. It occurs when a vessel supplying blood to the brain is obstructed due to the narrow blood vessel or clogs with fatty deposits. Particularly, acute ischemic stroke (AIS) due to large vessel occlusion (LVO) poses a large cerebral tissue at risk and carries high morbidity and mortality of the patients [2]. The leading cause of LVO is cardio-embolism due to atrial fibrillation. An increased incidence of atrial fibrillation with an aging population parallels the increase in LVO related AIS [3]. Besides applying clot-dissolving drugs and tissue plasminogen activator to treat AIS patients and restore blood flow in the blocked brain regions, endovascular mechanical thrombectomy (EMT) is also recommended to treat some AIS patients with LVO reduce morbidity and mortality [4]. Several multicenter clinical trials have recently reported improved treatment outcomes in AIS patients with LVO [5, 6]. In a study of operating EMT 6 to 24 hours after stroke with a mismatch between deficit and infarct, the results showed the improved outcome in patients treated with EMT plus standard care compared to patients who received standard care alone [5]. Additionally, similar improvement was also reported in another study conducted by EMT in stroke patients at 6 to 16 hours [6]. These trials are based on the estimation of cerebral infarct core and the existence of salvageable radiological brain tissue “at-risk” for infarction. Case selection biases in these trials are also widely debated. Additionally, the objective clinical or radiological correlation for region-specific AIS-related outcomes after such interventions is still unknown and needs an investigation in further studies.

Due to the lack of accurate clinical markers to accurately stratify AIS patients who can or cannot benefit from EMT to date [7], identifying new clinical markers that highly associate with the efficacy of EMT plays a critical role in quickly restoring the peripheral blood supply in a short period, which can help minimize the amount of brain tissue injury or risk of permanent tissue damage among the AIS patients who can benefit from EMT. Principles of currently used imaging software to select patients for EMT depends on arterial input function (AIF) and venous output function (VOF) to provide an estimate regarding cerebral blood flow (CBF), cerebral blood volume (CBV), mean transit time (MTT) and time to peak (TTP or Tmax) computed from computed tomography perfusion (CTP) images [8]. Such algorithms that depend on initial contrast flow in a major cerebral arterial system and outflow through a major venous channel are unable to capture microcirculatory dynamics of contrast flow through the brain parenchyma. Using CTP derived CBV to estimate ischemic core or AIS volume is often inaccurate as compared to that estimated using magnetic resonance imaging (MRI) based diffusion-weighted imaging (DWI) [9]. Some researchers highlighted that these inconsistencies are partially contributed due to high noise sensitivity in deconvolution-based on singular value decomposition. However, even though newer techniques using Bayesian method are robust as compared to SVD, they are still missing a significant infarct volume visible on the MRI images [10]. In a recent study, our group showed that clinical outcome is highly associated with the final cerebral infarct volume estimated using DWI sequences of post-intervention MRI [11]. Hence, although CTP has significant clinical advantages over MRI in AIS diagnosis due to its highly efficient and wide accessibility, it requires developing new imaging markers that have significantly increased prediction accuracy or high association to AIS patients’ clinical outcome with LVO.

Despite the improved imaging technology, qualitative image assessment of AIS severity or status using the radiologists-rated cerebral blood flow, cerebral blood volume, mean transit time, and time to peak [12] has limitations, including lack of quantitative assessment and large inter-reader variability [13]. To overcome these limitations, identifying and developing new quantitative image markers has been attracting broad research interests in the medical image informatics field [14]. In particular, the recently developed radiomics concept has proven that radiographic images (i.e., CT and MRI) depict useful image phenotype features that are highly associated with genomic biomarkers [15] and have the potential to predict disease prognosis [16]. Thus, based on radiomics concept and previous research focus in developing computer-aided detection and diagnosis (CAD) schemes of medical images [17 –21], including a CAD scheme equipped with an interactive graphical user interface (GUI) to detect and quantify the severity of aneurysmal subarachnoid hemorrhage patients using brain CT images [22], we aim to investigate the feasibility of developing new quantitative image markers computed from CTP images at an early diagnosis stage to predict AIS prognosis in this study. For this purpose, we developed a new CAD scheme with several novel image processing algorithms to segment the contrast agent enhanced blood volumes in bilateral cerebral hemispheres of the brain, generate cumulative blood flow curves and then compute asymmetrical blood flow related features in two brain hemispheres. Then, image markers based on the best single feature and machine learning (ML) models fused with multi-features are developed and tested to predict clinical benefit or outcome in a group of AIS patients undergoing EMT for LVO. The details of the study design and experimental data analysis results are presented in the following sections of this article.

2 Materials and methods

2.1 Image dataset

A de-identified retrospective dataset of pre-intervention CTP images of 31 AIS patients due to LVO was obtained from the Department of Neurology at the University of Oklahoma Health Sciences Center (OUHSC). Based on the current clinical standard of the Modified Rankin Scale (mRS) [23], the primary treatment outcomes of AIS patients are categorized into seven scales (from 0 to 6). Figure 1(a) summarizes the distribution of mRS among these 31 patients. Due to the small dataset, we divided the patients into two classes of good (favorable) and poor (unfavorable) prognosis based on mRS, as shown in Fig. 1(b). Class-0 includes 16 cases in which mRS ranges from 0–3, representing no symptoms to moderate neurological disability (requiring some help but also walking unassisted). While class-1 includes 15 cases ranging from 4–6 in mRS representing from moderately severe disability (unable to walk and attend to bodily needs without assistance) to dead. This study aimed to develop and apply new quantitative image markers to classify AIS cases into these two classes.

Fig. 1

Distribution of patients based on the Modified Rankin Scale (mRS). (a) Separated by individual mRS, (b) Separated by mRS into two classes: [‘class-0’: 0–3]; [‘class-1’: 4–6].

During AIS diagnosis and treatment in OUHSC, each patient is pre-assessed radiologically for their eligibility to undergo EMT. Specifically, multiple CT image scans are conducted during the image acquisition, including initial non-contrast CT of the head, CT angiogram, and sequential CTP scans. In CTP image acquisition, a rapid intravenous infusion of 40 ml of Isovue-370 contrast agent is administered. This contrast agent gradually enters and passes through the bloodstream and vessels, which visually distinguishes blood flow from other brain structures. Thus, during the CTP image acquisition process, the dynamic flow of contrast agent (i.e., wash-in and wash-out patterns) is used to capture the total blood amount and velocity of blood flow through different brain regions.

2.2 Image pre-processing

Among the clinical cases, there is variability in the number of CTP scanning sequences (i.e., ranging from 28 to 89), scanning range (i.e., whole brain or only the targeted volume of interest), resulting in a different number of image slices in one scanning sequence (i.e., ranging from 8 to 23), and image slice thickness (i.e., 2.5 or 5.0 mm). The pixel spacing parameters for all the cases are 0.488×0.488 mm for length and width respectively. Besides, some cases can use a one-directional CTP scanning protocol, while others use a two-directional scanning protocol. In a one-directional scan, when completing one scanning sequence of the targeted brain section, the CT machine pauses and returns to the starting point to perform the following scanning sequence (i.e., always from top to bottom or vice versa). Whereas in use of the two-directional protocol, the CT machine cyclically captures image sequences continuously without any break (i.e., scanning from top to bottom and then reversely scanning from bottom to top). Regarding image reconstruction algorithm, we used GE CT scanners, automatic settings of filtered back projection (FBP) with a standard convolution kernel and the filter selected as head. Then, the image pixel intensity values are converted into standard Hounsfield Unit (HU) value based on the dicom rescaling parameters such as slope and intercept.

In order to accommodate the acquired CTP image sequence irregularities mentioned above and make all cases comparable to each other, it requires CAD scheme to automatically organize the images retrieved from the clinical picture archiving and communication system (PACS) database by identifying the number of CTP image scanning sequences and adaptively labeling each image slice to a specific indexed brain location in the correct image sequence. Since the head is held fixed during image acquisition, the degree of similarity between the matched brain sections during the adjacent scanning sequences is higher. Thus, the CAD scheme uses a simple dice-similarity-based approach to identify two parameters (scan-type), one-directional or two-directional scanning protocols, and the number of the detected unique brain matching sections in different scanning sequences) for each case.

Our dataset found that the maximum number of scanned image slices in one scanning sequence for all patients is≤23 images. Thus, for each case, the CAD scheme selects the first 50 images retrieved from each CTP case to initially identify scan-type and match images in two adjacent scanning sequences. For this purpose, the CAD scheme first performs a rough brain segmentation using thresholding to exclude the skull region. If the number of connected regions in an image slice is more than one, only the most significant area is included, while others are discarded. Next, the CAD scheme uses the dice similarity coefficient to compute the similarity among these segmented areas in the images to identify the first unique CTP scanning sequence and its best matching pair. Thus, determining both the scan-type and the actual number of CTP image slices in each scanning sequence. Figure 2 illustrates this CAD process. For example, for one case using a one-directional scanning protocol with a total slice number of x, the CAD scheme requires detecting its scan-type and the number of x slices in one scanning sequence. Thus, slice (1) matches slice (x + 1) and continues (as shown in the middle row of Fig. 2). The bottom row of Fig. 2 shows that if the case uses a two-directional scanning protocol with a total slice number of x + 2 in one scanning sequence, CAD will identify image matching in a different order (i.e., slice x + 2 matches slice x + 3, and so forth). Once detecting the scan-type and the number of image slices in one sequence using the initial set of 50 image slices, CAD maps the results to the rest of all images in one case. Thus, the total number of CTP image scanning sequences of each case is computed using Equation (1), and all images of every series are labeled to a specific indexed brain location. $Total number of CTP sequences = \frac{Total number of images in one case}{number of image slices in one CTP sequence}$ (1)

Fig. 2

A sample illustration of the proposed dice-similarity based approach identifying the parameters (scan-type and number of unique brain indices).

2.3 Image segmentation

The CAD scheme implements a novel segmentation algorithm using image markers and mapping techniques. For each unique CTP image sequence, the CAD scheme identifies three perfusion markers, namely, global minimum (g_m), local minimum (l_m), and maximum peak (m_p). Like the method discussed in image pre-processing, CAD first performs an initial brain segmentation using thresholding to identify the largest connected intracranial brain region. A line plot depicting the initially segmented areas of each slice is constructed. Then, the slice with the largest area is considered as m_p, the slice with the smallest area is regarded as g_m, and the slice with the smallest area in the opposite direction concerning the m_p slice and g_m slice is marked as l_m. Next, the fine-tuning procedure is applied to segment final brain areas beginning with the m_p slice as the initial starting point. This slice’s actual brain is usually the most significant, single-connected component enclosed within the skull. Therefore, a precise segmentation without any leakages can be attained on this slice using a thresholding-based segmentation. In this way, the fine-tuning segmentation process continues towards either left or right direction, applying a consecutive mapping technique. This segmentation method uses the prior slice segmentation result to act as a reference for limiting boundaries to avoid segmentation leakages and identify the multiple-connected brain regions if existing. The limiting boundary criterion prevents segmentation leakage, which can be corrected by applying morphological dilation. Each region is examined and constrained to enclose within the limiting boundary for inclusion in the current slice segmentation in multiple connected regions. In summary, this segmentation process initiates at m_p slice for each sequence. It continues using the steps mentioned above until either a g_m slice or l_m slice is reached in both directions covering all the image slices.

Since this study primarily focuses on understanding and analyzing the asymmetry of blood flow between the left and right hemispheres of the brain, which is a critical image feature used by neuro-radiologists to assess the efficacy of EMT, CAD scheme splits each segmented brain slice into two parts of the left and right hemispheres. We also design and implement an interactive graphical user interface (GUI) of the CAD scheme with multiple visual-aid tools and functionalities, as shown in Fig. 3. Additionally, if GUI shows a slight tilt in brain image orientation during image acquisition, a functioning tool has been added in the GUI to request a CAD scheme to rotate images and correct image orientation. Thus, the CAD scheme can correctly separate the brain’s left and right hemispheres for image feature computation and data analysis.

Fig. 3

Picture of the implemented interactive graphical user interface (GUI) of the CAD scheme, which includes two image windows showing the original CT image slice (left) and the segmented brain area (right), and multiple operating functionalities and parameter assignment boxes on both left and right column.

2.4 Image feature computation

The CAD scheme first computes blood volume in the left and right hemispheres from the blood profile image of each CTP slice. CAD computes the following image features after grouping the computed blood values per each unique sequence. First, the blood volume (V_Blood) in one CTP slice is computed using Equation (2), $V_{Blood} (slice) = N B \cdot P (slice) \times P_{L} \times P_{B} \times S_{T} (slice)$ (2)

where N_B . P is the number of segmented blood pixels, P_L and P_B represent pixel length and breadth, while S_T represents slice thickness, respectively. A summation of V_Blood for all slices for one CTP scanning sequence (or a series) is computed using Equation (3), $V_{Blood} (series) = \sum_{slic e_{k} = 1}^{m} (N_{B \cdot P} (slic e_{k}) \times P_{L} \times P_{B} \times S_{T} (slic e_{k}))$ (3)

where m = number of slices in the series. However, V_Blood per series can be represented as two independent terms, voxel count (VC_term) and voxel parameter (VP_term), as explained from Equations (4 –6). $V C_{term} = N_{B \cdot P} [slic e_{k = 1,} slic e_{k = 2} \dots slic e_{k = m}]$ (4) $V P_{term} = P_{L} \times P_{B} \times S_{T} [slic e_{k = 1,} slic e_{k = 2} \dots slic e_{k = m}]$ (5) $V_{Blood} (series) = V C_{term} \times {(V P_{term})}^{'}$ (6)

where VP_term is an array of constant values for all the series. VC_term changes for each series depending upon the blood profile. Thus, V_Blood is directly proportional to VC_term. Throughout the rest of this article, we represent the summation of the VC_term as V_Blood, as indicated in Equation (7). $V_{Blood} (series) \infty (V C_{term} = \sum_{slic e_{k = 1}}^{m} N_{B \cdot P} (slic e_{k}))$ (7)

Second, to detect the trend of blood supply in two brain hemispheres over time, the CAD scheme computes the cumulative volume of blood (V_Cumulative) as shown in Equation (8), $V_{Cumulative} = [V_{s_{1}}, V_{s_{1} + s_{2}}, V_{s_{1} + s_{2} + s_{3},} \dots, V_{s_{1} + s_{2} + s_{3} \dots + s_{n - 1} + s_{n}}]$ (8)

where V_{s₁+s₂+s₃+…+s_n-1+s_n} is the summation of V_Blood between 1^st to n^th CTP scanning series, and n indicates the number of images in a unique series. This cumulative volume of blood (V_Cumulative) is calculated independently for both left (V_{Cumulative _ L}) and right (V_{Cumulative _ R}) hemispheres from their respective blood profile images using the steps as mentioned earlier from Equations (2 –8).

Third, due to variation in operator settings during CTP image acquisition, the scan brain regions and scanning duration vary among the patients. If a line plot is mapped between the number of unique series (x-axis: n) and cumulative volume of blood in a hemisphere (y-axis: [V_{Cumulative _ L} or V_{Cumulative _ R}]) for all the cases, the scales will not be compatible. To address this, we performed a case-based normalization of n, V_{Cumulative _ L}, and V_{Cumulative _ R,} as shown in Equations (9 –11), to scale or normalize the computed feature values between 0 and 1. $\hat{n} = \frac{n - min (n)}{max (n) - min (n)}$ (9) $Mi n_{val} = min (V_{Cumulative_L,} V_{Cumulative_R}); Ma x_{val} = max (V_{Cumulative_L}, V_{Cumulative_R})$ (10) ${\hat{V}}_{Cumulative_L} = \frac{V_{Cumulative_L} - Min_val}{Max_val - Min_val}; {\hat{V}}_{Cumulative_R} = \frac{V_{Cumulative_R} - Min_Val}{Max_Val - Min_Val}$ (11)

2.5 Asymmetrical blood flow pattern analysis

Since acute ischemic stroke (AIS) usually occurs in one hemisphere of the brain with LVO, which blocks the respective primary arterial blood flow. As a result, blood flow velocity in two hemispheres of the brain is different, making the dynamic flow (wash-in and wash-out of contrast agent) faster in the healthy hemisphere without LVO than the diseased hemisphere with LVO. Thus, detecting and quantifying asymmetrical blood flow patterns is our focus to identify new image markers to predict AIS prognosis. Specifically, as shown in Fig. 4, we divide the timed image sequences into three equal phases: initial, intermediate, and final. For each phase, the corresponding section of the line segments (V_{Cumulative _ L} and V_{Cumulative _ R}) from the line plots is utilized to calculate intermediate slopes (m_AB) using a linear regression method, where A is the right or left hemisphere, B is the phase. For example, m_L2 is the slope of the left hemisphere in the 2^nd/intermediate phase. Additionally, we subtracted the values of V_{Cumulative _ L} and V_{Cumulative _ R} between the left and right hemispheres to construct an absolute difference in cumulative volumes (|V_{Cumulative _ D}|) followed by normalization to yield ${\hat{V}}_{Cumulative_D}$ . Then, ${\hat{V}}_{Cumulative_D}$ curve is also divided into three equal phases similar to the method discussed earlier to compute their respective intermediate difference slopes (m_DB), where D represents the difference line, and B is the phase (as shown in Fig. 4). Finally, the absolute cumulative disparity in the total blood volume between both hemispheres after CTP image acquisition is computed using the last value of the arrays ( ${\hat{V}}_{Cumulative_R (end),} and {\hat{V}}_{Cumulative_L (end)}$ ) as shown in Equation (12). $abs_d {ifference}_{Total} = abs ({\hat{V}}_{Cumulative_R (end)} - {\hat{V}}_{Cumulative_L (end)})$ (12)

Fig. 4

A sample illustration of sectoring cumulative volume of blood line plot into three equal phases and computing corresponding intermediate slopes for left and right hemisphere.

2.6 Machine learning classifier model

Using the absolute cumulative disparity value computed at the end of the blood line plot (abs _ difference_Total) and other slope-based features computed in 3 phases (as shown in Fig. 4), we test several models to develop image markers to classify cases into two mRS classes of good and poor prognosis. Each model or marker generates classification scores ranging from 0 and 1. The higher score represents the higher likelihood of the case having a poor prognosis (‘class-1’: mRS = 3–6). Table 1 lists 3 independent models.

Model-I only uses one image feature of the absolute cumulative disparity value (abs_difference _Total) to simulate what neuro-radiologists do to predict patient prognosis with the quantitative data.

Model-II is built using features computed separately from two blood flow curves of the left and right hemispheres of the brain $({\hat{V}}_{Cumulative_L,} {\hat{V}}_{Cumulative_R})$ in three phases and the difference in cumulative blood volume(abs_difference_Total).

Model-III is built using features computed from one subtracted blood flow curve between the left and right hemispheres of the brain $({\hat{V}}_{Cumulative_D})$ in three phases and the differences in cumulative blood volume (abs_difference_Total), which emphasizes the bilateral asymmetry of the computed image features between the brain’s left and right hemispheres.

Table 1
List of features included in each type of ML model

Model Feature Pool

Model-I abs_difference_Total

Model-II m_L1,m_L2,m_L3,m_R1,m_R2,m_R3, and abs_difference_Total

Model-III m_D1,m_D2,m_D3, and abs_difference_Total

Model	Feature Pool
Model-I	abs_difference_Total
Model-II	m_L1,m_L2,m_L3,m_R1,m_R2,m_R3, and abs_difference_Total
Model-III	m_D1,m_D2,m_D3, and abs_difference_Total

In order to build multi-feature fusion models (Model-II and Model-III), we select two well-known supervised machine learning (ML) architectures, namely, support vector machine (SVM) and K-nearest neighborhood (KNN). Based on our previous experience of applying SVM and KNN in developing CAD schemes of medical images [24, 25], a polynomial kernel is used in the SVM model, and K = 5 (neighbors) is applied in the KNN model. To build the optimal ML models, the following three steps are used. First, a feature-wise normalization is performed to transform each feature’s values to a scale from 0 to 1. Second, a principal component analysis (PCA) method is applied to generate a new feature vector with a variance rate of 95% applied to reduce the redundancy of the image features. Third, due to the small dataset, a leave-one-case-out (LOCO) based cross-validation method is adopted to train and evaluate each ML model to maximize the number of training samples and avoid case partition bias [26]. In this way, each of the 31 cases in our dataset will be independently tested by the model trained using the other 30 cases in 31 training iterations.

To evaluate the performance of each classification model, we used the following two steps. First, a receiver operating characteristic curve (ROC) is constructed from the classification scores. The area under the ROC curve (AUC) is computed and used as an assessment index to evaluate and compare each model’s performance to classify between two mRS classes. Second, we apply an operating threshold on the classification scores (T = 0.5) to divide all testing cases into two mRS classes (score≤0.5: ‘Class-0’; score > 0.5: ‘Class-1’). From the classification results, several confusion matrices corresponding to different models are generated, which are used to compute various performance indices (i.e., classification accuracy, positive predictive value (PPV), negative predictive value (NPV), sensitivity, and specificity).

In summary, Fig. 5 shows a complete flow diagram of each step in our CAD scheme, which includes image pre-processing and data analysis pipelines using the ML model. All programs used in the CAD scheme and GUI tool (shown in Fig. 3) are coded using MATLAB R2019b package and libraries. Whereas the ROC curve and the AUC are computed using a maximum likelihood-based ROC curve fitting program (ROCKIT, http://www-radiology.uchicago.edu/krl/, University of Chicago), which is publicly available and widely used in the radiology and medical imaging informatics field.

Fig. 5

A detailed flow diagram of each step of the proposed CAD scheme.

3 Results

Figure 6 shows an example of the matched brain CTP slices in the whole scan of 28 sequences, which depict the change in the visibility of the contrast agent in the blood over the CTP image acquisition time. Looking at these 28 images from the top left to the bottom right in a left to right fashion, one can notice that the visibility of blood contrast is more dominant in the right hemisphere at the early phase of CTP scans as compared to the left hemisphere (wash-in). As the CTP acquisition continues, the blood contrast in the right hemisphere appears to drain completely (wash-out) first, whereas the contrast is still visible in the left hemisphere in the images of later scanning sequences. This example shows that our CAD scheme automatically detects the amount of contrast-filled blood volume over time using the cumulative amount of blood to quantify the contrast agent variation between the brain’s left and right hemispheres.

Fig. 6

From top-left to bottom-right: A sample brain index over CTP acquisition time depicting the variation in blood flow between the left and right hemisphere.

Figure 7 demonstrates the segmentation results of a sample CTP brain series for an individual patient. Unlike a simple segmentation scheme that targets to identify either a single connected cerebral region or uses the skull as the limiting boundary may fail to achieve accurate results as it may miss certain regions or cause possible leakages in some other scenarios, our CAD scheme successfully detects the multiple connected regions as shown in the 2^nd and 3^rd images of Fig. 7. The reason for generating such a result is that our CAD scheme uses 3 image markers as guides to set a protocol with the limiting boundary criterion for brain regions depicting each image slice, thereby identifying all the true multiple connected brain regions resulting in more sophisticated and accurate segmentation results.

Fig. 7

Illustration of proposed segmentation scheme using image markers and consecutive mapping technique for a sample brain series.

Figure 8 illustrates and compares the cumulative blood volume between the left and right hemispheres in two cases. In case (a), there is a clear big difference in the cumulative blood flow and volume between two hemispheres throughout the CTP image acquisition period, which indicates the presence of major LVO. Thus, applying EMT to remove the blood clot and resume blood supply can balance blood in both hemispheres since this patient can benefit from the EMT and thus receive a good clinical outcome (in class-0 of mRS). However, case (b) has a relatively small difference in cumulative blood flow or transit time in both hemispheres. Thus, the poor prognosis’s underlying reason is not primarily caused by LVO or an unbalanced blood supply. The clinical result shows this patient does not benefit from EMT and is classified into class-1 of mRS in this study.

Fig. 8

Comparison between two cumulative blood flow curves in left and right hemispheres of the brain, where case (a) is classified to ‘class-0’ and case (b) is classified to ‘class-1’ of mRS.

By analyzing all 31 testing cases in our dataset, Table 2 shows and compares the number of input features generated using the PCA algorithm and used to train ML models and the classification performance (AUC values) of 5 models. The corresponding ROC curves are presented in Fig. 9. The results show that Model-III built using the features computed from the subtracted blood flow curves (related to the transit time for contrast agent wash-in and wash-out) between two hemispheres of the brain produces the highest AUC values as compared to Model-I and Model-II. Using KNN and SVM ML methods, Model-III.2 (KNN) and Model-III.1 (SVM) yield AUC = 0.878±0.077 and 0.846±0.078, respectively.

Table 2

Summary of the number of PCA features used in various ML models and their corresponding classification performance in terms of AUC and standard deviation (STD)

Model	Number of Features	AUC±STD
Model-I	1	0.772±0.084
Model-II.1: SVM	5	0.746±0.089
Model-II.2: KNN	5	0.607±0.103
Model-III.1: SVM	4	0.846±0.078
Model-III.2: KNN	4	0.878±0.077

Fig. 9

Comparison of various ROC curves generated using 5 models to classify between two mRS classes.

Table 3 illustrates 5 confusion matrices of 5 models. Based on these matrices, a set of the computed performance indices, including accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), are summarized in Table 4. From these tables, one can observe that, like the AUC value assessment (Table 2), the highest classification accuracy is obtained for Model-III, in which Model-III.2 (KNN) achieves the highest prediction accuracy of 90.3%.

Table 3

Summary of confusion matrices of various ML models to classify between mRS classes

Model	Actual	Predicted
		Positive	Negative
Model-I:	Positive	11	5
	Negative	4	11
Model-II.1: SVM	Positive	13	3
	Negative	6	9
Model-II.2: KNN	Positive	10	6
	Negative	5	10
Model-III.1: SVM	Positive	13	3
	Negative	4	11
Model-III.2: KNN	Positive	16	0
	Negative	3	12

Table 4

Summary of several performance indices for various ML models

Model	Accuracy (%)	Sensitivity	Specificity	PPV	NPV
Model-I:	70.97	0.69	0.73	0.73	0.69
Model-II.1: SVM	70.97	0.81	0.60	0.68	0.75
Model-II.2: KNN	64.52	0.63	0.67	0.67	0.63
Model-III.1: SVM	77.42	0.81	0.73	0.76	0.79
Model-III.2: KNN	90.32	1.00	0.80	0.84	1.00

4 Discussion

In this study, we investigate the feasibility of developing and applying new quantitative image markers or ML models to predict AIS patients’ prognosis at an early stage using pre-intervention brain CTP images. The study has several unique characteristics and contributions. First, we apply several novel image processing algorithms to develop a new CAD scheme that can be applied to real clinical images with varying imaging scanning conditions. Based on our literature search, no similar CAD schemes are available to date. Our CAD scheme can automatically perform the following tasks including (1) organizing and matching all CTP image slices in the correct order of scanning sequences, (2) segmenting brain volume and contrast-enhanced blood volume in all CTP image slices, (3) generating two cumulative blood flow curve diagrams of left and right hemispheres of the brain, and (4) computing image features related to the blood transit time and velocity in 3 normalized phases of CTP scanning sequences. A unique interactive GUI is designed and used to increase reliability and user confidence in the CAD scheme (as shown in Fig. 3). As a result, any possible image processing errors (i.e., brain or blood volume segmentation errors) can be visually observed and corrected automatically by performing the pre-installed correction functions in the GUI or manually by the user’s hand drawing or new boundary condition setting. Although we did not test the GUI functions in this study, the similar GUI tool developed in our previous study [22] has been tested and used by the clinical researchers in the Department of Neurology of OUHSC to provide a quantified percentage of blood leakage volume in aneurysmal subarachnoid hemorrhage (aSAH) patients using brain CT images, and then predict clinical outcome of the patients as reported in our previous publications [27 –29].

Second, due to the potential presence of unilateral blood clots (LVO) in AIS patients, the transit time and velocity of the blood contrast flow rate (wash-in and wash-out) may vary between two hemispheres of the brain. Thus, asymmetrical blood flow rate or pattern in two hemispheres of the brain provides a potentially helpful image marker to predict AIS prognosis. Instead of subjective assessment of asymmetrical blood flow rate or transit time by the neuro-radiologists, which is qualitative and has large inter-reader variability, our CAD scheme computes several quantitative features to assess asymmetrical blood flow rate and patterns. Our data analysis results demonstrate that using the absolute cumulative disparity value has higher discriminatory power to predict AIS patients’ prognosis with AUC = 0.772±0.084 (Table 2), which is significantly higher than random guess (AUC = 0.5). Our GUI tool (Fig. 3) also provides clinicians (i.e., neuro-radiologists) a visual-aid tool to examine or monitor transit time or velocity of blood flow and the final difference of the cumulative amount of blood flow in the left and right hemispheres of the brain.

Third, another advantage of developing a CAD scheme is to compute multiple features. Then, ML methods can be applied to select optimal features and build multi-feature fusion models aiming to achieve an increased prediction performance than using a single optimal feature or marker. In this study, we investigate two sets of features. As shown in Fig. 4, one includes cumulative blood flow related features computed separately from the brain’s left and right hemispheres. In contrast, another one includes features computed from the subtracted cumulative blood flow curve of two hemispheres. The study results (Table 2) demonstrate that multi-feature fusion models yield significantly higher prediction performance than using a single image feature or marker. Model-III also yields significantly higher performance than model-II for using both SVM and KNN learning methods. It indicates that using the absolute difference curve of blood flow between two hemispheres carries more discriminatory information or power to train ML models than using the features computed separately from two hemispheres. It also reduces the number of features (Table 2), which can help improve ML models’ robustness. Additionally, performing paired t-test showed that 4 out of 6 analyses have a significant statistical performance (p < 0.05) amongst the multi-feature fusion models. For instance, the best performing Model-III.2 (in terms of accuracy and AUC), has statistically significant performance as compared to Model_II.1 and Model-III.1. More details related to the p-values computed between the various ML methods are summarized in Table 5.

Table 5

Comparison of p-value between various ML models

P-value	Model-II.1	Model-II.2	Model-III.1	Model-III.2
Model-I	0.238689	0.20936	0.218763	0.234386
Model-II.1		0.009978	0.382433	0.012067
Model-II.2			2E–05	0.25409
Model-III.1				2.65E–07

Fourth, many radiomics studies have recently reported that radiographic images (i.e., CT) contain useful phenotype image features, which are highly associated with the prognosis of cancer patients [15, 16]. This study demonstrates the feasibility of identifying and applying radiographic image features or markers computed from brain CTP images to phenotype AIS patients and predict their prognosis potentially. This study supports and expands the radiomics concept to more broad clinical application fields. The success of our approach to developing new quantitative image markers or prediction models will eventually provide clinicians a new decision-making supporting tool to assist them more accurately stratifying AIS patients for choosing and applying optimal treatment methods (i.e., EMT) at an early stage aiming to reduce patients’ mortality and morbidity rates.

Despite the encouraging results, we also recognize that this study has several limitations. First, only a small dataset is used in this study. The distribution of cases belonging to each mRS class (0 to 6) is also not uniform, which restricts us to divide the dataset into only two classes to represent good and poor prognosis. Due to the small dataset, the computed AUC values have a relatively big standard deviation (Table 2), indicating a relatively lower confidence level. Additionally, a small dataset may not well represent the general AIS population in the real clinical environment. Thus, our CAD scheme’s performance and robustness and image marker or ML models need to be further tested using larger and independent study cohorts in future studies. Second, the simple case normalization to compensate for the different numbers of the CTP scanning sequences and division of cumulative blood volumes into three equal phases may not be optimal. A more dynamic or adaptive division of cumulative blood volume phases should be investigated and compared in future studies. Third, to fully use the radiomics concept, more image features need to be explored and computed to identify more discriminatory information to improve ML models’ performance and robustness. Fourth, in this study, segmentation depends on thresholding and labeling algorithms to identify brain and blood volume. Deep learning technology has been currently applied to segment the targeted regions of interest in medical images [30 –32]. Using the deep learning method may help achieve higher accuracy in the segmentation of brain and blood volume using CTP images in the future. Last, we cannot compare this new software tool with other software tools developed by other researchers or commercial companies due to the lack of access. However, we believe that this study will contribute to this research and development field due to our unique image processing and data analysis approaches and the good preliminary testing results (AUC = 0.878±0.077). We will conduct a more comprehensive study using a large new dataset to compare our new software tool with other state-of-art tools once they are either publicly available or implemented in our medical center in the future.

In conclusion, this is a preliminary and proof-of-concept study to develop new quantitative image markers to classify AIS patients based on mRS severity using a set of bilateral asymmetrical image features computed of CTP images between the left and right hemispheres of the brain. The study demonstrates promising results when applying the CAD scheme and ML model to a set of diverse clinical cases with different mRS distribution and varying CTP imaging scanning protocols. Based on the foundation built in this study, a new research effort can follow to validate these quantitative image markers further and conduct prospective clinical studies in the future.

Footnotes

Acknowledgments

This study is supported in part by research grants R01CA197150 and P20GM135009 from the National Institutes of Health, USA.

Declaration of competing interest

The authors declare that they have no competing interests.

References

National Center for Health Statistics. Health, United States, 2017:With special feature on mortality. Hyattsville, MD. 2018. https://www.cdc.gov/nchs/data/hus/hus17.pdf.

Malhotra

, Gornbein

and Saver

J.L.

, Ischemic strokes due to large-vessel occlusions contribute disproportionately to stroke-related dependence and death: a review, Front Neurol 8 (2017), 651.

Ray

and Keyrouz

, Management of anticoagulant-related intracranial hemorrhage: an evidence-based review, Critical Care 18 (2014), 223.

Adams

H.P.

, Adams

R.J.

, Brott

, et al., Guidelines for the early management of patients with ischemic stroke: a scientific statement from the Stroke Council of the American Stroke Association, Stroke 34 (2003), 1056–1083.

Nogueira

R.G.

, Jadhav

A.P.

, Haussen

D.C.

, et al., Thrombectomy 6 to 24 hours after stroke with a mismatch between deficit and infarct, N Engl J Med 378 (2018), 11–21.

Albers

G.W.

, Marks

M.P.

, Kemp

, et al., Thrombectomy for stroke at 6 to 16 hours with selection by perfusion imaging, N Engl J Med 378 (2018), 708–718.

French

B.R.

, Boddepalli

R.S.

and Govindarajan

, Acute ischemic stroke: current status and future directions, Mo Med 113(6) (2016), 480–486.

Hoeffner

H.G.

, Case

, Jain

, et al., Cerebral perfusion CT: technique and clinical applications, Radiology 231(3) (2004), 632–644.

Copen

W.A.

, Morais

L.T.

, Wu

, et al., In Acute Stroke, Can CT perfusion-derived cerebral blood volume maps substitute for diffusion-weighted imaging in identifying the ischemic core? PloS one 10(7) (2015), e0133566.

10.

Sakai

, Delman

B.N.

, Fifi

J.T.

, et al., Estimation of ischemic core volume using computed tomographic perfusion, Stroke 49 (2018), 2345–2352.

11.

O’Connor

K.P.

, Hathidara

M.Y.

, Danala

, et al., Predicting clinical outcome after mechanical thrombectomy: The GADIS (gender, age, diabetes mellitus history, infarct volume, and current smoker [corrected]) score, World Neurosurgery 134 (2020), e1130–e1142.

12.

Wintermark

, Sincic

, Sridhar

and Chien

J.D.

, Cerebral perfusion CT: technique and clinical application, J Neuroradiol 35(5) (2008), 253–260.

13.

Zhao

, Tan

, Bell

D.J.

, et al., Exploring intra- and inter-reader variability in uni-dimensional, bi-dimentional, and volumetric measuremnts of solid tumors on CT scans reconstructed at different slice intervals, Euro J Radiol 82(6) (2013), 959–968.

14.

Kim

TY.

, Son

and Kim

K.G.

, The recent progress in quantitative medical image analysis for computer aided diagnosis systems, Healthc Inform Res 17(3) (2011), 143–149.

15.

Aerts

H.J.W.L.

, Velazquez

E.R.

, Leijenaar

R.T.H.

, et al., Decoding tumor phenotype by noninvasive imaging using a quantitative radiomics approach, Nat Commun 5 (2014), 4006.

16.

Liu

, Gu

, Hu

, et al., Applying a radiomics-based strategy to preoperatively predict lymph node metastasis in the resectable pancreatic ductal adenocarcinoma, J Xray Sci Technol 28(6) (2020), 1113–1121.

17.

Tan

, Zheng

, Leader

J.K.

and Gur

, Association between changes in mammographic image features and risk for near-term breast cancer development, IEEE Trans Med Imaging 35 (2016), 1719–1728.

18.

Aghaei

, Tan

, Hollingsworth

A.B.

and Zheng

, Applying a new quantitative global breast MRI feature analysis scheme to assess tumor response to chemotherapy, J Magn Reson Imaging 44(5) (2016), 1099–1106.

19.

Danala

, Thai

, Gunderson

, et al., Applying quantitative CT image feature analysis to predict response of ovarian cancer patients to chemotherapy, Acad Radiol 24 (2017), 1233–1239.

20.

Danala

, Patel

, Aghaei

, et al., Classification of breast masses using a computer-aided diagnosis scheme of contrast enhanced digital mammograms, Ann Biomed Eng 46(9) (2018), 1419–1431.

21.

Gong

, Liu

, Jiang

, et al., Fusion of quantitative imaging features and serum biomarkers to improve performance of computer-aided diagnosis scheme for lung cancer: A preliminary study, Med Phys 45 (2018), 5472–5481.

22.

Aghaei

, Ross

S.R.

, Wang

, et al., Implementation of a computer-aided detection tool for quantification of intracranial radiologic markers on brain CT images, Proc SPIE 10138 (2017), 1013805.

23.

Broderick

J.P.

, Adeoye

and Elm

, The evolution of the Modified Rankin Scale and its use in future stoke trials, Stroke 48(7) (2017), 2007–2012.

24.

Heidari

, Khuzani

A.Z.

, Hollingsworth

A.B.

, et al., Prediction of breast cancer risk using a machine learning approach embedded with a locality preserving projection algorithm, Phys Med Biol 63 (2018), 035020.

25.

Chen

, Khuzani

A.Z.

, Hollingsworth

A.B.

, et al., Applying a new quantitative image analysis scheme based on global mammographic features to assist diagnosis of breast cancer, Comput Methods Programs in Biomed 179 (2019), 104995.

26.

and Doi

, Reduction of bias and variance of evaluation of computer-aided diagnostic schemes, Med Phys 33 (2006), 868–875.

27.

Nouh

C.D.

, Ray

, Xu

, et al., Quantitative analysis of stress-induced hyperglycemia and intracranial blood volumes for predicting mortality after intracerebral hemorrhage, Transl Stroke Res (2022), doi: 10.1007/s12975-022-00985-x.

28.

Ray

, Ross

S.R.

, Danala

, et al., Systemic response of coated-platelet and peripheral blood inflammatory cell indices after aneurysmal subarachnoid hemorrhage and long-term clinical outcome, J Crit Care 52 (2019), 1–9.

29.

Danala

, Desai

, Ray

, et al., Applying quantitative radiographic image markers to predict clinical complications after aneurysmal subarachnoid hemorrhage: A pilot study, Ann Biomed Eng (2022), DOI: 10.1007/s10439-022-02926-z.

30.

Nishio

, Koyasu

, Noguchi

, et al., Automatic detection of acute ischemic stroke using non-contrast computed tomography and two-stage deep learning model, Comput Methods Programs Biomed 196 (2020), 105711.

31.

Heidari

, Mirniaharikandehei

, Khuzani

A.Z.

, et al., Improving performance of CNN to predict likelihood of COVID-19 using chest X-ray images with preprocessing algorithms, Int J Med Inform 144 (2020), 104284.

32.

Tan

, Zhou

, Li

, et al., Automated vessel segmentation in lung CT and CTA images via deep neural networks, J Xray Sci Technol 29(6) (2021), 1123–1137.

Developing new quantitative CT image markers to predict prognosis of acute ischemic stroke patients

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSIONS:

Keywords

1 Introduction

2 Materials and methods

2.1 Image dataset

Table 1 List of features included in each type of ML model Model Feature Pool Model-I abs_difference Total Model-II mL1,mL2,mL3,mR1,mR2,mR3, and abs_difference Total Model-III mD1,mD2,mD3, and abs_difference Total

Footnotes

Acknowledgments

Declaration of competing interest

References

Table 1
List of features included in each type of ML model

Model Feature Pool

Model-I abs_difference_Total

Model-II m_L1,m_L2,m_L3,m_R1,m_R2,m_R3, and abs_difference_Total

Model-III m_D1,m_D2,m_D3, and abs_difference_Total