How Can Radiomics Help the Clinical Management of Patients with Acute Ischemic Stroke?

1. Introduction

Acute ischemic stroke (AIS) is the loss of neurological function due to a sudden reduction in cerebral blood flow and is a leading cause of disability and death worldwide [1]. According to the WHO, stroke is the second leading cause of disability and death worldwide after ischemic heart disease, accounting for 11% of all deaths. With improvements in the quality of life and longevity of patients, the incidence of stroke is expected to continue to rise in the coming years [2]. More than 80% of strokes are ischemic in origin [3]. Recent advances in the management of these patients have focused more on their treatment, where endovascular therapies have radically changed the prognosis of these patients [4,5,6]. In recent years, numerous studies have explored the many diagnostic aspects of ischemic stroke. However, since the introduction of cerebral perfusion studies more than a decade ago, technological advances with clinical applicability have been somewhat stagnant [7].

Today, medicine is immersed in a digital world. Specialties such as radiology are experiencing an unprecedented boom. In recent years, with the advent of artificial intelligence, applications such as radiomics have emerged to harness this technological revolution and improve diagnostic techniques in the search for personalized clinical management of patients. Since the coining of the term radiomics in 2012 [8], articles on this topic have become increasingly common.

Radiomics is a radiological discipline based on the idea that images are more than just pictures and that there is a lot of data that is invisible to the human eye, but which we can quantify and analyze by looking for data patterns associated with certain pathologies. Initially, radiomics was mainly focused on the field of oncology [9]. All cancer patients have at least one staging CT scan, so the most commonly used imaging technique for feature extraction in these studies was CT [10]. The aims of these studies were very diverse. From the determination of prognostic factors, estimation of overall survival, prediction of treatment response... The use of radiomics for gene determination (also called radiogenomics) has also been developed in pathologies such as brain tumours [11], lung cancer [12] and abdominal tumours [13]. In recent years, the use of radiomics has even been extended to other imaging modalities such as MRI, positron emission tomography (PET) based on both CT and MRI, and ultrasound. The aims of these articles were similar to those of the CT-based articles.

In the field of neuroradiology, radiomics features obtained from magnetic resonance imaging (MRI) were used most frequently in the articles. In these articles, radiomics also focused mainly on oncological patients. The objectives varied from differentiation of high-grade gliomas from brain metastases, prediction of EGFR mutations, prognostic prediction, determination of overall survival... In recent years, the use of radiomics has extended to brain pathologies other than tumours. There are articles on the use of radiomic features for prognostic prediction in acute ischemic stroke (AIS), for predicting degenerative diseases, or even for predicting aneurysm rupture.

The aim of this review is to summarize published articles on the performance of radiomics in patients with AIS, and to consider the wide range of possibilities offered by radiomics in these patients. We are currently at a key point in the evolution of the radiology specialty, as has been the case with other technological advances such as diffusion MRI, helical CT or, more recently, dual-energy CT. We need to be at the forefront of this technological evolution so that our patients can benefit.

2. Radiomics

2.1. Radiomics Workflow

The radiomics workflow consists of 5 consecutive steps (Figure 1):

• Obtaining images: Radiomics can be applied to all types of medical imaging. There are published articles on the performance of radiomics in ultrasound, Non-Contrast Enhancement-CT (NCECT), Computed Tomography Angiography (CTA), MRI and PET. In this step, there can be a great deal of heterogeneity when using images acquired in different hospitals or even on different machines within the same hospital. Therefore, the acquired images are subjected to a standardization process in which we try to correct this source of heterogeneity [8].
• Pre-processing: The quality of the images can be improved by using the pre-processing tools. In this step, some image filters are used to reduce noise. The aim is to increase the predictive power of the classifiers [14].
• Segmentation: In this step, the region of interest is selected in the radiological image. The segmentation of this region can be done in three ways: manual, semi-automatic and automatic. Manual segmentation is the gold standard and the most commonly used method in the studies reviewed in this article. The main advantage of this model is the intervention of an expert radiologist in its performance. The main disadvantage of this model is the time required to manually segment the entire region of interest [15]. Automatic segmentation is based on automatic detection of the region of interest without human intervention. Finally, semi-automatic segmentation is performed under the supervision of an expert radiologist who can edit an initial automatic pre-segmentation. The advantage of this method is the speed of the segmentation and the fact that the human component remains [16]. With today's increasingly sophisticated segmentation software, semi-automatic 3D segmentation of an area of interest can be performed quickly and comfortably for the radiologist.
• Feature extraction and classification: There is a wealth of numerical data that can be extracted from medical images, known as radiomic features. There are several classes of radiomic features: Histogram features (grey level mean, maximum, minimum, variance and percentiles), texture features (absolute gradient, grey level co-occurrence matrix -GLCM-, grey level run length matrix -GLRLM-, grey level size zone matrix -GLSZM-, grey level distance zone matrix -GLDZM-, Neighborhood Grey Level Difference Matrix -NGTDM-, Neighborhood Grey Level Dependence Matrix -NGLDM-), model-based features, transform-based features (Fourier, Garbor, Wavelet) and shape-based features (geometric properties of ROIs) [14]. These numerical data are classified by automatic classifiers. These classifiers are capable of recognizing different groups of patterns depending on the objective we set for them. In this way, the classifiers learn the data patterns with a training cohort. Then, with a test cohort, these classifiers are able to classify the data we provide into the pattern that most closely resembles it. In addition to numerical data, automatic classifiers can be used to input clinical data to search for a combined model.
• Data analysis

2.2. Radiomics in AIS

Articles analyzing the performance of a radiomic model in neurological pathology have become increasingly frequent. The most frequently published articles are about neurological tumour pathology. However, in recent years there has been an increase in the publication of articles about non-tumour neurological pathology.

Regarding the application of radiomics in acute cerebral ischemic pathology, there has been an exponential increase in the publication of articles since 2018 (Figure 2). The objectives of these articles on radiomics in acute ischemic stroke are prognostic prediction of AIS patients, detection of early ischemic stroke, prediction of outcome of revascularization treatment, prediction of complications, prediction of the etiology of AIS, prediction of time since stroke and differentiation between hemorrhage or contrast extravasation in patients undergoing mechanical thrombectomy (Figure 3).

2.3. Prognostic Prediction

The use of radiomics for prognostic prediction of AIS patients is the most commonly published. Avery et al. in their study of 829 patients used Radiomic Features (RF) obtained from CTA images of patients with anterior circulation Large Vessel Occlusion (LVO). They used masks of the middle cerebral artery (MCA) territory to extract the RF. They used four different input sets: Radiomics (only RF of MCA territory), Radiomics + Treatment (RF, mTICI score and intravenous thrombolysis), Clinical + Treatment (clinical variables and treatment) and Combined (RF, treatment and clinical variables). The clinical variables included were sex, age and admission NIHSS. The best results were very similar between the clinical+treatment model and the combined model. The area under the curve (AUC) for long-term outcome prediction was 0.81 and 0.82 for the clinical+treatment and combined models, respectively. The AUC for predicting discharge outcome were 0.82 and 0.81 for the clinical+treatment and combined models, respectively. In the internal independent and external test performances, the best results of the different inputs were very similar. There was no significant difference between the best results of the radiomics only model, the radiomics + treatment model, the clinical + treatment model and the combined model. The authors concluded that RF could help with patients where the collection of clinical information is limited [18].

Other studies of some prognostic predictions have found similar results. Cui et al. obtained an AUC of 0.821, 0.968, 0.983 and 0.896 (for mRS 1, 2, 3 and 4, respectively) in predicting mRS 90 days after acute ischemic stroke [19]. Wang et al. also predicted the mRS scale after AIS in 598 patients. They obtained 402 radiomic features based on MRI. The AUC of the combined model was similar (AUC of 0.80) [20]. Gerbasi et al. obtained an AUC of 0.85 in predicting long-term functional outcome using a combined model with RF obtained from MRI [21]. Guo Y et al. in their study of 56 patients used clinical information such as age, NIHSS, hypertension, diabetes, atrial fibrillation, mRS... and RF obtained from dynamic susceptibility contrast perfusion-weighted images (DSC-PWI) to predict functional outcome. Their best results for predicting mRS 2 were AUC of 0.884 and Acc of 0.864. For the prediction of mRS 4 they achieved an AUC of 0.908 and an Acc of 0.821 [22]. Guo Y et al. also investigated functional recovery three months after stroke and achieved similar results (best AUC of 0.971) [23]. Jiang L et al. obtained similar results (AUC of 0.862 with the combined clinic-radiomics nomogram) in predicting outcome (mRS >2) in patients with AIS using DWI-based radiomics in 1716 patients [24]. Li et al. in 102 patients obtained an AUC of 0.88 and 0.79 in the training and validation sets for predicting poor prognosis using a combined model with clinical data and MRI-based radiomics [25]. Ramos et al. used a combined model based on CTA images. They extracted 1260 features from 3279 patients (MR CLEAN registry). They obtained a better AUC with the combined model in predicting TICI>=2b (0.61) than with the clinical model alone and concluded that the combination of radiomic features and clinical data improves the prediction of good reperfusion. In the prediction of the mRS scale, they did not obtain significant differences between the combined model and the clinical model [26]. Li et al. also obtained significant results in good prognostic prediction after mechanical thrombectomy in AIS patients (AUC of 0.945 and 0.920 in training and validation cohorts). They extracted 1396 radiomic features from MRI images of 260 AIS patients [27]. Quan et al. also predicted prognostic factors in AIS patients. They developed an MRI-based radiomics model to predict adverse outcome after AIS (mRS scale >=2). They obtained an AUC of 0.926 and 0.864 in training and validation cohorts, respectively, with a combined model. They concluded that MRI radiomic features can predict unfavorable clinical outcome in AIS patients [28]. Tang et al. also predicted favorable clinical outcome in AIS patients after thrombolysis with similar results in short-term clinical prediction [29]. Tolhuisen et al. also attempted to predict clinical outcome (mRS score 0-2 at 90 days) in AIS patients using MRI-based radiomics. They obtained 0.88 and 0.81 in the training and validation cohorts. They concluded that radiomics can provide additional information in predicting functional outcome in AIS patients [30]. Yu et al. also developed a radiomics model to predict clinical outcome (mRS scale <=2 or >2) in AIS patients. They obtained similar results (AUC of 0.902, accuracy of 0.831). They also concluded that radiomics models based on MRI imaging can predict clinical outcome in AIS patients [31]. Zhou et al. also developed a radiomics model based on MRI imaging of 522 AIS patients to predict clinical outcome (mRS scale >=2 or >2 at 6 months). They also obtained similar results (AUC of 0.868 and 0.890 in training and validation cohorts with the combined model). They concluded that the combined model outperformed individual clinical or radiomic models, with a satisfactory result in predicting clinical outcome in AIS patients [32].

2.4. Detection of Ischemic Stroke

In the detection of ischemic stroke, articles focus on the early detection of ischemic lesions and the prediction of recurrence. The group of Guan Y et al. showed in their study with 56 patients and 1301 RF that there is a correlation between RF and AIS detection (best Acc of 0.7748) [33]. Guo Y et al. also investigated the detection of ischemic stroke (IS) and the prediction of NIHSS and other outcomes using RF obtained from DSC-PWI in 88 patients. The best AUC they obtained was 0.925 for stroke detection. For the prediction of NIHSS and other prognostic predictions, they obtained an AUC of 0.853 and 0.828, respectively [34]. Zhang et al. also tried to predict early brainstem infarction using an NCECT-based radiomics model. They obtained an AUC of 0.99 and 0.91 in the internal and external validation cohorts, respectively. They concluded that NCECT-based radiomics can detect early brainstem infarction and may be very useful in clinical practice [35].

On the other hand, Zhang R et al. used MRI-based radiomics to predict the ischemic penumbra in their study of 241 patients. They extracted 896 radiomic features from the DWI high signal area using manual segmentation. The AUC obtained was 0.92, sensitivity, specificity and accuracy were 0.93, 0.75 and 0.82 respectively in the training cohort. In the validation cohort the AUC, sensitivity, specificity and accuracy were 0.90, 0.88, 0.74 and 0.80 respectively. They concluded that a radiomics model based on MRI images can identify the ischemic penumbra in AIS patients [36].

In ischemic stroke recurrence prediction, Su et al. developed a combined model with some clinical factors (age, sex, number of chronic ischemic lesions, routine laboratory test, carotid ultrasound, cardiovascular risk factors and medical treatment) and CT-based radiomics to predict future ischemic stroke in patients with silent lacunar infarction. They obtained a C-index of 0.786 and 0.714 in the training and validation cohorts, respectively. They concluded that radiomic features can provide important information for predicting future ischemic stroke in patients with silent lacunar infarction [37]. Tang et al. also developed a radiomics signature based on MRI features of patients with symptomatic intracranial atherosclerotic stenosis (SICAS) to predict stroke recurrence. They obtained an AUC with the combined model of 0.899 and 0.803 in the training and validation cohorts, respectively. They concluded that radiomic features are useful for predicting stroke recurrence in patients with SICAS [38]. Wang et al. also predicted 1-year ischemic stroke recurrence in 1003 AIS patients based on MRI images. They obtained similar results [39].

2.5. Treatment Predictions

Regarding radiomic models predicting the outcome of reperfusion treatments, most studies focus on predicting the first-pass effect or the TICI scale after thrombectomy. Only one study focuses on predicting recanalization after ateplase treatment.

To predict first-pass recanalization, Hofmeister et al. performed clot segmentation in their study of 156 patients. They used RF to identify patients with first-pass recanalization by aspiration. They also predicted the total number of passes required to achieve good recanalization with a mechanical thrombectomy device (MTB). They achieved an AUC of 0.88 for predicting patients with first attempt recanalization. They concluded that clot-based radiomics can predict an MTB strategy in patients with AIS [40]. Sarioglu et al. attempted to predict first attempt recanalization in AIS patients using clot-based radiomics features obtained from NCECT. They achieved an accuracy of 0.83 using a combined model with radiomic features and clinical data (ASPECTS scale and gender). They also concluded that clot-based radiomic features can help to estimate successful recanalization in AIS patients [41]. Patel et al. also developed a NCECT/CTA-based radiomic model to predict first-pass effect in AIS patients with similar results (AUC of 0.832) [42].

In recanalization prediction, Zhang et al. attempted to predict successful recanalization (TI-CI>=2c) using MRI-based radiomics in 141 AIS patients. They extracted 321 radiomic features from the ischemic lesion and performed a manual segmentation. They obtained an AUC of 0.7442, a sensitivity of 0.7556 and a specificity of 0.7675. They concluded that MRI-based radiomics may be useful in predicting patient response to thrombectomy [43]. Xiong et al. also attempted to predict successful recanalization (TICI >=2b) in AIS patients undergoing stent retrieval. They improved the results of Zhang et al. with an AUC of 0.860 and 0.849 in the training and validation cohorts with the combined model. They also conclude that the radiomics model can successfully predict recanalization after stent retrieval [44].

Van Voorst et al. developed a clot-based radiomics model using CT images of 699 patients to predict TICI>=2b, first attempt reperfusion, reperfusion within three attempts and mRS scale <=2. They obtained similar results with the clinical, radiomics and morphological features models, but found that thrombus texture was independently associated with reperfusion OR 0.85 (95% CI 0.72 to 0.99) [45].

Finally, Qiu et al. predicted recanalization after ateplase treatment in 67 AIS patients with ICA/M1 MCA segment thrombus. They obtained an AUC of 0.85. They concluded that clot-based radiomics was more predictive of recanalization with ateplase than other classic thrombus imaging features (length, volume and permeability) [46].

2.6. Prediction of Complications after AIS

There are some studies that have used RF to predict some complications in patients with AIS. Fu B et al. obtained an AUC of 0.96 in predicting malignant cerebral oedema (MCE) using a combined model with RF from NECT of patients with AIS [47]. Jiang et al. also predicted oedema after acute ischemic stroke using MRI-based radiomics in 389 patients, with similar results [48]. Wen et al. developed a radiomic model to predict malignant acute middle cerebral artery infarction (mMCAi) based on CT radiomic features (NCECT and CTA). They obtained an AUC of 0.917 and 0.913 in the training and validation cohorts, respectively. They concluded that radiomic features could be an instrumental tool to predict the risk of mMCAi [49].

On the other hand, Meng et al. attempted to predict hemorrhage transformation in AIS patients based on MRI radiomics features. They extracted 5400 radiomic features from 71 AIS patients. They performed manual segmentation. They obtained an AUC of 911 and an accuracy of 0.894 in predicting hemorrhage transformation with a combined model. They concluded that a combined model with MRI-based radiomics can provide important information to physicians in the diagnosis of AIS patients [50]. Xie et al. also developed a CT-based radiomics model to predict hemorrhagic transformation in AIS patients. They achieved similar results (AUC of 0.845-0.750). They also concluded that CT-based radiomics could help predict hemorrhagic transformation [51].

Finally, Liu et al. developed a radiomics model based on NCECT to predict hemorrhage expansion in patients with thrombolysis/thrombectomy related hemorrhagic transformation. They extracted 1691 features from a manually segmented region of interest in 104 patients. They obtained an AUC of 0.91 and 0.87 in a training and validation set. Sensitivity and specificity were 0.83-0.60 and 0.89-0.78 respectively. They concluded that CT-based radiomics is valuable in predicting hemorrhage extension [52].

2.7. Etiology Prediction

In the etiology prediction of AIS using radiomics models, Chen Y et al. tried to differentiate between cardioembolic and atherosclerotic etiology in patients with acute ischemic stroke. They included 82 patients and they obtained RF from CTA images. They did a manual segmentation of the embolic region and they extracted 116 RF. In their results, the best AUC was 0.9018 and the best Acc was 0.8929. They concluded that radiomics can effectively predict the subtype of ischemic stroke and help doctors to correctly manage these patients [53]. Jiang J et al. also studied with 403 AIS patients a clot-based model to identify cardioembolic (CE) stroke before MTB. They achieved an AUC of 0.838 for predicting CE stroke. They concluded that clot-based RF can reliably predict CE strokes [54].

2.8. Time Since Stroke Prediction

Chen Y et al. attempted to estimate the time since stroke onset (TSS). They manually outlined the clot on CTA images of 221 patients and obtained 944 RFs. They divided the patients into two different groups: <4.5 hours and >4.5 hours since stroke onset. They analyzed the performance of the radiomics model and the combined model (radiomics + clinical data). The AUC was 0.803 for the combined model and 0.813-0.803 for the combined model (training and test cohorts). The clinical data used were age, diabetes and atrial fibrillation. They concluded that there was no significant difference between the radiomics and combined models. They also concluded that radiomics based on RF obtained from CTA images can estimate TSS in patients with AIS [55]. Yao et al. also developed a CT-based radiomic model to classify the onset time of basal ganglia infarcts. They obtained the radiomic features from CT images of 316 patients. They classified the patients into two groups (>4.5h and <4.5h). In the training cohort, they obtained an AUC of 0.982, a sensitivity of 0.929 and a specificity of 0.959. In the validation cohort the results were very similar (0.974, 0.951 and 0.961). They concluded that radiomics can be a useful tool for predicting time since stroke in basal ganglia infarction [56]. Wen et al. also developed a CT-based radiomics model to predict infarct onset time in patients with MCAO. They obtained similar results with the combined model (AUC of 0.808-0.833) [57]. Finally, Zhang et al. also developed an MRI-based radiomics model to predict TSS with similar results (AUC of 0.754, accuracy of 0.788). They concluded that radiomics based on MRI imaging could help to guide thrombolysis decisions in patients with unknown stroke onset [58].

2.9. Differentiation Hemorrhage from Iodinated Contrast Extravasation after Thrombectomy

Chen X et al., in their very interesting study of 101 patients, tried to differentiate intracranial contrast extravasation from intraparenchymal hemorrhage (IPH) after mechanical thrombectomy (MT). They used RF from non-enhanced CT (NECT) images after MT. They used manual segmentation and extracted 1316 RF from 166 intraparenchymal areas of hyperattenuation. They obtained very interesting results, with an AUC between 0.848 and 0.826 (training and validation cohorts). The accuracy (ACC), sensitivity (S), specificity (Spec), negative predictive value (NPV) and positive predictive value (PPV) were 0.776, 0.767, 0.789, 0.682 and 0.852 respectively. They concluded that radiomics can effectively differentiate IPH from intracranial contrast extravasation after MT [59]. Yuan Ma et al. also used a CT-based radiomics model to predict bleeding from iodinated extravasation after thrombectomy in AIS patients. A total of 100 patients were included in the study. They performed a manual segmentation method and obtained 1316 radiomic features from NCECT. They obtained an AUC of 0.972 and 0.926 in the training and validation cohorts, respectively, using a combined model. They concluded that the combined nomogram was effective in differentiating between hemorrhage and iodinated contrast extravasation [60].

4. Conclusions

Radiomic models can provide valuable information for the management of patients with AIS. The results obtained with radiomic models are very good and encouraging. The wide variety of clinical situations in which radiomics provides valuable information in patients with AIS suggests that we may be looking at a model that will change the paradigm of clinical management of patients with AIS. Further prospective studies are needed to investigate the utility of radiomics in the mentioned fields of knowledge and others that may appear in the future.