Rapid Assessment of Acute Ischemic Stroke by Computed Tomography Using Deep Convolutional Neural Networks

Introduction

According to World Health Organization statistics, stroke is the most common form of cerebrovascular disease among adults, is one of the leading causes of death, and the major cause of permanent disability globally. Every two seconds, someone suffers a stroke, which results in approximately 6.24 million deaths worldwide each year [1]. Stroke patients are at high risk of recurrence, which increases from 3% (after 30 days) to 40% (after 10 years) [2]. Additionally, the Global Burden of Disease reports that up to 11 million stroke cases occur in young adults aged 20~64 years [3]. Stroke is a brain injury that induces physical impacts, such as limb impairment, dementia, aphasia, and cognitive abnormalities [4, 5], which impair patients’ independence. The required expenditures for medical resources and family burdens are enormous. Therefore, timely stroke recognition and treatment are urgently needed to reduce brain tissue injury as much as possible.

Stroke caused by disruption of blood supply to the brain can be classified into either rupture or blockage of blood vessels, namely, hemorrhagic and ischemic stroke, respectively [6]. Ischemic stroke is by far the most common type (up to 80%) worldwide [7]. The main cause is large vessel occlusion due to either an arterial or cardiac embolism [8]. The primary goal of critical treatment in the acute phase of ischemic stroke is to recover the supply of blood perfusion. Good clinical outcomes were correlated with early vessel recanalization [9, 10]. The correct use of treatment highly depends on an accurate diagnosis of acute ischemic stroke; therefore, a rapid and reliable imaging examination is necessary for subsequent treatment.

Non-contrast computed tomography (NCCT) and magnetic resonance imaging (MRI) are routinely used to diagnose ischemic stroke [11]. NCCT is a fast and readily available diagnostic tool for emergency situations and is used to detect cerebral hemorrhage (a contraindication to thrombolytic therapy) or other mimics of acute stroke. Diffusion-weighted imaging is a highly sensitive MRI technique that provides reliable information on acute cerebral ischemia [12]. Compared to NCCT, MRI is less accessible and more limited by patient intolerance, contraindications, and a long examination duration. Consequently, NCCT remains the first-line imaging modality for identifying ischemia extent in the brain parenchyma [13]. Nevertheless, these findings are present in fewer than 50% of cases [14]. Erroneous interpretations, such as sulcal effacement and inconsistent scanning protocols, limit the sensitivity of NCCT. In the literature, NCCT has a sensitivity of 57~71% within 24 h and only 12% in the first 3 h [15, 16].

With the development of image processing, computer-aided diagnosis (CAD) systems based on quantitative image features and machine learning techniques can achieve better efficiency and accuracy [17, 18], and the use of CAD in diagnosing acute ischemic stroke by NCCT can meet clinical needs. CAD systems are quantitative, efficient, and consistent; therefore, CAD systems have been proposed to improve diagnostic quality. For example, textural features are used to evaluate liver cirrhosis at an early stage based on ultrasound images [19]. Shape features are used to automatically detect pulmonary nodules in CT imaging [20] and detect masses and microcalcifications on mammograms [21]. Previous CAD systems used handcrafted shape, intensity, and textural features that were manipulated by human experts; however, selecting and implementing relevant features are complicated and may be biased. Deep convolutional neural networks (DCNNs) are a better approach to automatically generate and combine image features. DCNNs with multiple convolutional layers can extract as much information as possible from images, which are automatically processed with several types of filters and learn to recognize features at differing spatial levels [22]. Prior literature showed that the performance of DCNN in diagnosing diabetic retinopathy and skin cancer outperformed specialized expertise with an area under the receiver operating characteristic curves (AUC) for accuracy of 0.96–0.99 [23–25]. In this study, DCNN as a CAD system was proposed to diagnose acute ischemic stroke on NCCT images to provide an efficient and accurate diagnostic suggestion for clinical use. Additionally, NCCT image datasets were collected from different scanners and institutions for performance validation to evaluate the generalizability of the trained model.

Materials and Methods

Results

The performances of the four DCNN models using the dataset from institution A are listed in Table 2. Ten-fold cross-validation was applied to evaluate how the model could be generalized to an independent dataset. With tenfold cross-validation, the dataset was first randomly partitioned into 10 equal folds. Then, 10 iterations of training and validation were performed, where in each iteration, nine folds were used for training and a different fold of data was held out for validation.

AlexNet achieved an accuracy of 97.12%, a sensitivity of 98.11%, and a specificity of 96.08%, which were better than those of the transferred AlexNet: 90.49%, 96.98%, and 83.73%, respectively. Regarding transferred Inception-v3 and transferred ResNet-101, the accuracy, sensitivity, and specificity were 94.62%, 93.98%, and 95.29%, and 95.49%, 94.36%, and 96.67%, respectively. With respect to the trade-offs between the sensitivity and specificity, the observed AUC of AlexNet was 0.9927, which was also higher than transferred AlexNet (0.9895), transferred Inception-v3 (0.9889), and transferred ResNet-101 (0.9897) (Fig. 5).

Table 3 shows the independent test results regarding the dataset from institution B. AlexNet showed 60.89% accuracy, 18.52% sensitivity, and 100% specificity. Transferred AlexNet achieved 81.77% accuracy, 62.04% sensitivity, and 100% specificity. Transferred Inception-v3 achieved 85.78% accuracy, 75% sensitivity, and 95.73% specificity. Transferred ResNet-101 achieved 80.89% accuracy, 61.11% sensitivity, and 99.15% specificity. With respect to the example of misclassified cases in Fig. 6, tissues suffering from ischemia have complex anatomical structures and the pathologic findings may confuse the DCNNs.

Discussion

NCCT is used to rapidly identify parenchymal findings of ischemic stroke in the acute stage. When ischemia occurs, cytotoxic edema (intracellular) in the affected parenchyma can lead to tissue hypoattenuation on NCCT. NCCT signs may be subtle to the human eye, especially in the hyperacute stage, and there is substantial interrater variability in diagnosing early ischemic changes, ranging from 20 to 69% sensitivity with a very restricted time window [33, 34]. The reliability of an imaging finding is a critical parameter that determines its clinical utility.

In addition, these early signs are not useful for distinguishing between an ischemic penumbra and a core infarction. The Alberta Stroke Program Early CT Score (ASPECTS), an objective method to evaluate the extent of ischemia and prognosis that is widely used, was developed to standardize the analysis of early ischemic signs with excellent interobserver reliability, but is too time-consuming for radiologists to perform in clinical settings [35]. The interobserver reproducibility of ASPECTS varies depending on the reader’s experience, the amount of reader training, stroke knowledge, and stroke onset-to-imaging time. Reader inexperience and a lack of rigorous ASPECTS training were highlighted as potential causes of suboptimal interobserver agreement. Yong et al. [36] reported limits of ASPECTS as a predictor, such as small hypodensities (lacunar infarct) and age-related white matter hypodensities. Prior reports described that ischemic stroke detection within 12 h varied with pretreatment NCCT ASPECTS, including the location of the arterial occlusion, the time of stroke onset, and image acquisition [37, 38]. Furthermore, reperfusion therapy in ischemic stroke increases the risk of intracerebral hemorrhage. Symptomatic intracerebral hemorrhage occurs in 2.4~10% of patients within 24~36 h after thrombolysis and is associated with an increased risk of subsequent death or disability [39], so the timely recognition of early ischemic changes on NCCT is critical for improving stroke outcomes.

In this study, all NCCT images in the stroke group were obtained within 6 h after symptom onset. Using DCNN on the stroke classification achieved higher than 90% accuracy in the model training and test using the institution A dataset, as shown in Table 2. Considering that AlexNet without transfer learning still achieved high accuracy, although more training time was necessary, the accuracy implied that the collected dataset was enough to establish the acute ischemic stroke classification model from a specific scanner. Comparing AlexNet and transferred AlexNet, the transferred parameters were learned from color natural images in ImageNet, which may not be optimized for grayscale NCCT images. Tested by a dataset from institution B (Table 3), although AlexNet trained from scratch had only 60% accuracy, the transferred DCNNs achieved substantial performance (80–85%) and have potential clinical use. The experimental results demonstrate that training DCNN from scratch can be customized for a specific scanner. For example, a system is only applied in a single institution or for a scanner. In another way, if a CAD is used for general purposes, the parameters learned from large datasets such as ImageNet would provide more consistent accuracy. The accuracy differences between a customized model and a general model were presented in the experiments. The results show that the highest accuracy and the highest generalizability cannot be obtained simultaneously. To us a DCNN for various scanners, we suggest training a customized model for each scanner. In clinical use, accuracy is the first priority rather than convenience.

In comparison with other CAD systems based on segmentation, edge detection, and feature extraction, previous studies only achieved accuracies and sensitivities of 80~95% [40–42]. RAPID, as an automated ASPECTS in acute MCA stroke, provides an easy, reproducible, and structured method for CT reading [43]. Rutczy’nska et al. [44] proposed a Gaussian mixture model of regional growth for finding the stroke region, but the method was only effective for large lesions and failed to effectively detect early stroke signs. Wu et al. [45] developed an image patch classification-based method to detect ischemic regions using a radiomics-based patch model in CT, and those designs were handcrafted features. Choosing features is time-consuming and operator dependent. The proposed model, which learns relevant features by the training process in DCNN, aids radiologists in focusing on the diagnosis of the ischemic area and subsequent treatment rather than a complicated feature design. Promising results would be helpful in providing diagnostic suggestions in an automatic and consistent way. In particular, the proposed CAD system equipped with an Intel i7-9700 processor, an NVIDIA GeForce GTX 1080 TI graphics card, and 64-GB system memory only spends 0.1 s to determine whether an NCCT slice has stroke. In seconds, a series of brain slices can be evaluated for clinical use. By means of parallelizing and optimizing the implementation, the proposed CAD system will accelerate its practicality. Compared to Food and Drug Administration (FDA)-approved products, StrokeViewer does not provide accuracy or detect lesions for acute ischemic stroke. The use of neural networks for the evaluation model demonstrates that the power and efficiency of DCNNs are undoubted.

This study has some limitations. Although a customized model can be made for a specific scanner, one generalized model is most convenient to use to handle datasets from various scanners. Collecting many datasets remains a challenge. In future studies, we may consider establishing a database with representative images from various scanners and patients. Size number is an issue only if as the number increases, diversity also increases. Increasing cases similar to the existing training set is not helpful. As long as the dataset is representative, the amount is less important. In addition, more radiologists with substantial clinical experience should be recruited, and whether suggestions from CAD results can improve their accuracy in diagnosing acute ischemic stroke could be explored. To date, the proposed AlexNet model and the transferred DCNN models have performed well in the retrospective diagnosis of acute ischemic stroke using NCCT images.

Conclusion

Based on the developed scheme with DCNN architecture, acute ischemic stroke in NCCT images can be accurately diagnosed with an AUC of > 0.98 for a specific scanner. The accuracy differences between a customized model and a general model were presented. In clinical use, the accuracy of a customized model is more important than its generalizability. Rapid triage and treatment in emergency settings can be achieved with CAD suggestions to radiologists, which will improve treatment quality and result in better patient care.

Funding

The authors received financial support from the Ministry of Science and Technology of Taiwan (MOST 108-2221-E-004-010-MY3), the University System of Taipei Joint Research Program (USTP-NTPU-TMU-107-01), and the Mackay Memorial Hospital (MMH-106-150).

Declarations