Ischemic stroke is a significant cause of death worldwide and has a heavy toll on death and disability (1). Therefore, the warning symptoms, clinical features, and prognostic evaluation around stroke have always been the subject of clinical and scientific research. Research shows that the initiating presentation of ischemic stroke is the occlusion of a blood vessel that impairs blood flow to a certain degree, leading to infarction of brain tissue in the part of the brain supplied by that blood vessel (2, 3). That means the state of cerebral blood flow has become a significant factor for the early warning and status assessment of stroke, and the early detection of abnormal blood flow is of great significance for timely treatment and excellent prognosis of patients.

To identify the presence of reduced regional blood flow, studies and physicians have combined diverse modalities of images with various analysis methods to detect abnormal states and identify hypoperfusion areas (HA) that may cause a stroke. In most imaging modes, perfusion images, such as dynamic susceptibility contrast perfusion-weighted imaging (DSC-PWI) and computed tomography perfusion (CTP), play a vital role in stroke analytics in clinical practice and trials due to their ability to evaluate cerebral blood flow state. When the contrast agent arrives at the ill-perfused tissue of the brain, the signal intensity values barely change since there is no or less propagation of the contrast agent to the damaged tissue (4). Thus, the time to the maximum tissue residual function (T_max) obtained from DSC-PWI, a highly commonly used parameter, has been used in clinical trials to identify the HA (5, 6). Generally, the region recognized from the condition T_max >6s is defined as HA (7). In addition, other single-modality images except for DSC-PWI, or in combination with it, can also provide much medical information on ischemic stroke. A study (8) shows that the mismatch between DSC-PWI and diffusion-weighted imaging (DWI) has been used to estimate the ischemic penumbra and provides a valuable tool in the clinical treatment of stroke, which helps guide the selection of the clinical therapeutic plan. Lu et al. (9) evaluated the volume of the ischemic penumbra using susceptibility-weighted imaging and mapping (SWIM) of patients with asymmetrical prominent cortical veins. Wang et al. (10) discussed the value of susceptibility-weighted imaging (SWI) in evaluating the ischemic penumbra of patients with acute cerebral ischemic stroke. Bhattacharjee et al. (11) verified that the quantitative assessment of the penumbra using the SWI-DWI mismatch ratio performs equivalently to the ASL, PWI-DWI mismatch ratio. Furthermore, continuously developed artificial intelligence models can interpret and analyze the manifestations of stroke (12–14). Although many previous studies have been committed to evaluating the HA from multimodal imaging manners, multidimensional analysis methods, and advanced artificial intelligence technology, there are few methods to analyze the image features themselves to discover the association between the image features and cerebral blood flow state.

Radiomics is an emerging methodology that quantifies high-dimensional features from imaging data and has been used to investigate tumor heterogeneity (15, 16) and for clinical decision support systems to improve treatment decision-making and accelerate advancements toward precision medicine in cancer (14, 17–21). Recently, only a tiny minority of studies have investigated the role of radiomics in identifying ischemic stroke lesions (22), evaluating prognostic biomarkers based on the penumbra (23), and predicting functional outcomes (24). However, these studies combined medical images with clinical text information to perform the above tasks but ignored the features themselves. Currently, few studies have explored the association between imaging characteristics in the temporal dimension of DSC-PWI and blood flow status in ischemic stroke. However, with abundant and distinct blood flow information in DSC-PWI data, it is possible to extract these features to explain the blood flow state.

As for classification tasks, machine learning models and neural networks have been widely used for a long time. However, each method has its rules and algorithms to perform tasks. For example, Logistic Regression (LR) (25–27) quantifies the coefficients of variables to predict a logit transformation of the probability of the presence of the event. Support Vector Machine (SVM) (28) learns an optimal hyperplane that separates the classes as widely as possible. SVM can also perform nonlinear classification using the “kernel” to map to higher dimensional feature space (29). Random Forest (RF) is created based on decision trees (DT). Their methods resemble human reasoning by representing hypotheses as sequential if-then. The AdaBoost algorithm (Ada) (30) corrects the misclassifications made by weak classifiers, and it is less susceptible to overfitting than most learning algorithms. Gradient Boosting Decision Tree (GBDT) adapts the boosting algorithm, and it uses the error rate of the previous iteration weak learner to update the weight of the training set (31). Besides, the k-nearest neighbor (KNN) is a non-parametric classification method that forms the k neighborhood for features (32). The Naive Bayes classifier (NB) is a simple probabilistic classifier based on Bayes's theorem under solid independence between components (33). In addition to numerous machine learning models, neural networks, such as Multilayer Perceptrons (MLP) and Convolutional Neural Networks (CNN) (34–36), are commonly used to perform classification tasks. Generally, a single model is usually selected for task execution in the current classification tasks. However, as there are more or fewer differences between the algorithms of different models, the comprehensive evaluation of the results through multiple models will increase credibility. Thus, if we can verify the performance of the image features of DSC-PWI data in identifying ischemic stroke through models with different preferences, it will undoubtedly improve the validation accuracy and enhance the persuasiveness.

The purpose of this paper mainly consists of two aspects. The first is to discover the image features hidden in DSC-PWI data that can accurately distinguish normal tissues from abnormal tissues. The second is to explore the changes in the classification task when the proportion of abnormal tissues is different.

Detailed materials and methods are introduced in the following. The procedures in this study include making HA and normal area (NA), computing radiomics features, selecting excellent features, and evaluating radiomics feature sets.

Results are provided in three parts, including extracted significant radiomics features, selected outstanding features, and the performance of the selected feature sets. The details are shown in the following.

An ischemic stroke is a vascular event characterized by reducing regional blood flow. Few studies explored the changes among DSC-PWI images in the time dimension, although the parameter T_max obtained from them was commonly used to discriminate HA and NA. Some studies (49–51) have shown that the time-intensity curve of HA in the DSC-PWI images of patients with ischemic stroke has a much smaller brightness decrease than the curve of NA (seen in Figure 10). Therefore, the data of DSC-PWI in the time dimension are correlated with the blood flow state of brain tissues to a certain extent. This study successfully extracted multi-level feature selection processing and the radiomics features distinguishing HA and NA from DSC-PWI. Of all the methods, the F_Lasso reached the best CS of 0.925, and the five indexes were mAcc of 0.958, mPre of 0.96, mAuc of 0.982, mF1 of 0.959, and mRecall of 0.96. Besides, we effectively verified the ability of these features to evaluate the ischemic area ratio in the brain. According to the results, with the increase of the proportion of ischemic tissue, the mAcc increased, while Pre stabilized and then decreased. And the best Pre and Acc can reach 0.888 and 0.863. In general, the radiomics features of 3D images in the time dimension of DSC-PWI have an optimistic ability to distinguish normal brain tissue from abnormal brain tissue and indicate the proportion of ischemic tissue in brain tissue.

This study used 13 feature selection methods with different preferences to obtain outstanding features. As a result, there were 128 excellent features selected from the original 65,800 radiomics features. Their time values are mainly concentrated at the initial moment (0–3), the stage through which the contrast agent passes (17–22), and a few features located at the end of the reaction (time >30) (seen in Figure 2C). The results indicate that the initial intensity of the tissue, as well as the amount of intensity change, and the time producing the change, are essential to distinguish between normal and abnormal tissues. Besides, the features of Shape, GLRLM, GLSZM, and NGTDM were missing in the 128 selected features (seen in Figure 2D). This means that the shape, gray of neighboring voxels, and length in the number of pixels with the same gray make little contribution to characterizing the changing of blood flow, and features in the other groups are significant. Among 13 feature sets, F_Lasso and F_Alpha in F_SSL achieved the best CS of 0.925 and 0.904, while F_FI performed worst, and F_SIF and F_STFperformed in the middle (seen in Figure 4). From the four attributes (p-value, R squared, Gain, and Gain ratio), the p-value of F_FI is more significant than the others. In contrast, the R squared, Gain, and the Gain ratio are less than the others, suggesting that the effect of features extracted by FI may not be ideal. In addition, the feature selection methods (CMIMI, JMI, MIFS, MIM, and MRMR) in the FI mainly select features based on the conditions of information entropy, redundancy, and similarity between radiomics features to obtain feature sets. Among the 65,800 original radiomics features, there may be a large number of features that meet the above screening conditions. However, in this study, the number of selected features of the feature selection method is limited to 10. Although this study ranks features according to their scores, the selected features may not be complete, resulting in the unsatisfactory performance of F_FI. This can be further improved and verified in subsequent experiments. Different from FI, other feature selection methods in SIF, STF, and SSL selected features through the linear relationship, contributions, and statistical scores between features and sample categories. When sorting features with scores, excellent features will be selected first, so it is reasonable that they got a relatively higher performance than FI. In detail, For the feature sets in F_SIF, the attributes of F_LS have less correlation and information than the others, and they got a matching result that the CS of F_Fisher and F_ReliefF are better than F_Lap. F_FS and F_TS in F_STF got similar attributes and achieved the closer CS of 0.874 and 0.865. For feature sets in F_SSL, although F_Lasso and F_Alpha obtained a minor R squared, Gain and Gain ratio than feature sets in F_STFand F_SIF, they achieved the best performance. For the long term, Lasso has been used to select excellent features and has been validated in the fields of classification (52–54), prediction (55, 56), and survival analysis (57, 58). In this study, Lasso got the best feature set F_Lasso to prove its competence in screening features. Thus, although the lower p-value, higher correlation, information gain, and information gain ratio can achieve a better classification result to a certain extent, they cannot be used as complex indicators to evaluate their effectiveness. The 13 F_method are a great deal of diversity, and these selected features are highly significant. No matter what selection method is used, they can obtain the characteristics of DSC-PWI from the aspects of intensity variation, drop time of intensity, initial state, and recovery state.

Furthermore, this study analyzed the classification ability of radiomics features in different proportions of ischemic lesions. With the increase of RT, the region of ischemic tissue increases, and the difference between features whose RT is above the set threshold and those of the opposite class decreases. When RT was <0.25, regardless of S = 3, 4, or 5, the Acc and Pre can reach >0.8. However, when RT was >0.25, the performance will decrease with the increase of RT. On the one hand, the decline of these two indexes may be due to the imbalance in the proportion of positive and negative samples when RT reaches 0.25. If sufficient data are available, in-depth reason analysis can be performed in the future. Nevertheless, these results demonstrate that the radiomics features can effectively distinguish normal tissue from ischemic tissue, provide support for the differentiation of volume proportion of ischemic lesions and provide information for clinical guidance.

In addition, ten models with different principles were used to verify the performance of selected features. The ten models included regression models (LR, NB), nonlinear classifiers (SVM, DT, RF), linear classifiers (KNN, DA), ensemble models (Ada, GBDT), and neural networks (NN). According to the classification results of these models, the classification effect of selected features can be verified comprehensively. Figures 4, 6–8 show little difference in the performance of the same feature set in different models. Still, there is a significant difference in the performance of different feature sets in the same model, and SVM, LR, NN, RF, and DA performed better than the others. For the classification of HA and NA, SVM performed best in almost all feature sets, with an mAUC of 0.929. In particular, the Auc, Pre, Acc, F1, and Recall of CSVM_Lasso were all >0.987. Using a nonlinear kernel 'RBF' in SVM, the nonlinear relationship between the selected radiomics features and the target (stroke tissue or not) can be found, thus obtaining accurate classification results. Besides, DA, RF, NB, LR, and NN also achieved satisfactory results. Regarding identifying the proportion of ischemic stroke, DA and SVM also performed better than the other models in all three situations, RT_2, RT_3, and RT_3. Although the performance of different models on the same feature set and the same situation had good consistency, SVM was a better choice in both evaluation tasks, classifying HA and NA and identifying the proportion of ischemic stroke. Besides, we used CS computed by the mean indexes (mAcc, mPre, mAuc, mF1, and mRecall) of the ten models as the benchmark for the evaluation to reasonably analyze their performance. Depending on the diversity features, the 13 F_method acquired different CS ranging from 0.624 to 0.925, FRT_3′, FRT_4′, and FRT_5′ got CS in [0.34,0.76], [0.40, 0.78] and [0.28,0.78], respectively. On the one hand, the strong robustness and applicability of the Lasso algorithm can be proved by the fact that, although the features extracted by the algorithm were slightly different under different RT values, the extracted features generally achieved stable performance. On the other hand, the selected radiomics features at different slices have little influence on the classification results, but the proportion of ischemic tissue does.

There are some limitations to this study. First, the size of the datasets is relatively small, and all data come from a single hospital, which may lead to biased results and a lack of generalizability. To address the limitation, we segmented hypoperfusion areas (HA) from DSC-PWI images and defined normal tissue in the symmetrical areas of HA as NA in making ROIs. This way, one group of HA and NA can be generated from one DSC-PWI image. This way, the double samples (160) can be obtained from 80 DSC-PWI images, and the positive and negative sample sizes are equal. The expanded balanced samples can help extract accurate features, and the sample imbalance can be reduced when classifying NA and HA. Besides, when evaluating the performance of the selected features in section Performance evaluation of the selected features, the tenfold cross-validation was performed to reduce the influence of sample size. The composite scores (CS) were computed to obtain reliable results. Second, the feature selection methods, optimal features, and learning models can be further optimized. This paper uses various existing learning models to verify the classification performance. Although the results have shown some features such as F_Lasso and F_Alpha had achieved excellent performance, the further optimization of the models, such as deep learning and transferred learning, can be regarded as one of the future works. The ischemia area ratio classification needs to be further improved. The results in this study do not mean that the models can be used alone for stroke treatment decision-making. Instead, it should be considered a support tool for stroke treatment guidance. We will validate our improved method's performance with more data before applying it to clinical trials in future work.

This study used prominent radiomics features extracted from 3D images in the DSC-PWI time series to explore their ability to classify HA and NA and recognize the proportion of ischemic lesions in brain tissue. The 13 F_method achieved the CS ranging from 0.624 to 0.925 in distinguishing HA from NA. The F_Lasso in the 13 F_method performed best with mAcc of 0.958, mPre of 0.96, mAuc of 0.982, mF1 of 0.959, and mRecall of 0.96. Besides, the 120 FRT_S′ reached the best CS of 0.786 in identifying the proportion of the ischemic region, and the best Acc and Pre reached 0.888 and 0.863, respectively. In general, the combination of various radiomics features accurately reflected the varying degrees of changes in cerebral blood flow in the initial state, the contrast agent response stage, and the recovery stage. For classifying the proportion of ischemic areas, the classification effect is relatively stable when RT is <0.25. Otherwise, when RT was >0.25, the accuracy will gradually decrease as its increases. Further future research should be conducted on excellent feature extraction, feature combination, model optimization, and comprehensive verification.

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author/s.

The studies involving human participants were reviewed and approved by Shanghai Fourth People's Hospital Affiliated to Tongji University School of Medicine, Shanghai 200434, China. The patients/participants provided their written informed consent to participate in this study. Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

YLu, WL, and YK conceived of the idea and encouraged YG to investigate and supervised the findings of this work. YLi provided the datasets. YG, YY, FC, MF, XZ, WQ, and LC developed the theory and performed the computations. YG, QL, XM, and YLi compared the analytical methods. YG, LL, and XZ analyzed and discussed results. YG, YLi, and YY wrote and corrected the manuscript with input from all authors. All authors discussed the results and contributed to the final manuscript.

The study was supported by the Stable Support Plan for Colleges and Universities in Shenzhen, China (SZWD2021010), the Scientific Research Fund of Liaoning Province, China (JL201919), and the National Natural Science Foundation of China (62071311).

Author LC was employed by Shenzhen Happy-Growing Intelligent CO., Ltd.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.