The value of CT-based radiomics in predicting hemorrhagic transformation in acute ischemic stroke patients without recanalization therapy

doi:10.21203/rs.3.rs-2967403/v1

Download PDF

Research Article

The value of CT-based radiomics in predicting hemorrhagic transformation in acute ischemic stroke patients without recanalization therapy

https://doi.org/10.21203/rs.3.rs-2967403/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

To investigate the clinical value of radiomics based on non-enhanced head CT in the prediction of hemorrhage transformation in acute ischemic stroke (AIS).

Materials and methods

The radiomic features of infarcted areas on non-enhanced CT images were extracted using ITK-SNAP. The Max-Relevance and Min-Redundancy (mRMR) and the least absolute shrinkage and selection operator (LASSO) were used to select features. The radiomics signature was then constructed by multiple logistic regression. The clinicoradiomics nomogram was constructed by combining radiomics signature and clinical characteristics. All predictive models were constructed in the training group, and these were verified in the validation group. All models were evaluated with the receiver operating characteristic (ROC) curve, calibration curve, and decision curve analysis (DCA).

Results

The radiomics signature was constructed by 10 radiomics features. The clinicoradiomics nomogram was constructed by combining radiomics signature and atrial fibrillation. The area under the ROC curve (AUCs) of the clinical model, radiomics signature, and clinicoradiomics nomogram for predicting hemorrhagic transformation in the training group were 0.64, 0.86, and 0.86, respectively. The AUCs of the clinical model, radiomics signature, and clinicoradiomics nomogram for predicting hemorrhagic transformation in the validation group were 0.63, 0.90, and 0.90, respectively. DCA curves showed that the radiomics signature performed well as well as the clinicoradiomics nomogram. DCA curve showed the clinical application value of radiomics signature is similar to that of clinicoradiomics nomogram.

Conclusion

Radiomics signature which was constructed without clinical characteristics can independently predict the hemorrhagic transformation of AIS well.

hemorrhagic transformation

acute ischemic stroke

recanalization

Acute ischemic stroke (AIS) has the characteristics of high incidence, high fatality rate, and high disability rate, which is one of the leading causes of death in the world [1]. Hemorrhagic transformation is one of the most common and worrying complications of acute ischemic stroke, with an incidence of 10% and 40% reported in the literature [2]. The infarcted brain tissue is prone to hemorrhage transformation, and cerebral hemorrhage can lead to serious deterioration of neurological function and even life-threatening [3], which has a significant impact on the prognosis of patients [4]. Most of the hemorrhagic transformation is caused by the natural course of AIS, and it can also be caused by thrombolysis, interventional thrombectomy, and other recanalization therapy [5]. Recanalization therapy such as thrombolysis and interventional thrombectomy is currently considered to be the most effective treatment for acute ischemic stroke, which has been proven to significantly improve the neurological impairment and prognosis of patients [6, 7]. However, because of its narrow treatment time window, not many patients can receive recanalization therapy. Therefore, for the vast majority of patients who have not received recanalization treatment, it is more important to predict the occurrence of hemorrhagic transformation in the early stage.

Early identification of patients with a high risk of hemorrhagic transformation will lead clinicians to reduce the dose of antithrombotic drugs such as aspirin tablets, adopt neutral treatment options, and conduct more frequent clinical evaluations and monitoring [8]. Furthermore, the re-examination time of cranial CT can be appropriately shortened and the frequency of cranial CT can be increased to detect the hemorrhagic transformation as soon as possible, which has guiding value for the subsequent adjustment of the treatment plan in clinical practice [9]. For patients with a lower risk of hemorrhagic transformation, clinicians can take more active treatment plans to appropriately reduce the frequency of clinical monitoring and cranial CT examination, so that patients can get more clinical benefits and reduce treatment costs. Therefore, the technical method that can accurately predict the hemorrhagic transformation after AIS patients is helpful for clinicians to make individual and accurate clinical treatment plans.

In recent years, radiologic features such as the high-density sign of the middle cerebral artery, low density of CT or abnormal diffusion-weighted imaging (DWI) signal on admission, and very low cerebral blood volume have been considered to predict the hemorrhagic transformation, may change the clinical management plan of patients, and help clinicians to make early prevention before the hemorrhagic transformation occurs [2]. However, the current research shows that it is not enough to rely on these radiologic features to predict hemorrhagic transformation, so it is necessary to explore a more accurate and objective model to predict the hemorrhagic transformation of AIS. It is also pointed out in the literature that the predictive model of hemorrhagic transformation also has the value of prognosis evaluation to a certain extent [10].

A variety of machine learning algorithms combined with medical images have been used in the diagnosis and prognosis of AIS. A multicenter study showed that the attention-gated U-Net deep learning algorithm with DWI and MRI perfusion as inputs could predict the final infarct volume, independent of the reperfusion state, and significantly overlap with the basic performance of the fluid-attenuated inversion recovery (FLAIR) sequence obtained 3–7 days after onset (Dice score, 0.53 × FLAIR 0.31–0.68) [11]. Machine learning algorithms, including regularized Logistic regression, linear support vector machine, and random forest, are superior to existing pre-treatment scoring methods in predicting the clinical outcome of patients with macrovascular occlusion undergoing thrombectomy. The AUC of the machine learning model is 0.85–0.86, while the AUC of the pre-treatment score is 0.71–0.77 [12]. Previous studies have found that the hemorrhagic transformation after AIS was related to factors of clinical such as hypertension, age, hyperglycemia, and stroke severity [13, 14]. The use of a machine learning algorithm to combine medical image quantitative information with clinical information can effectively predict hemorrhagic transformation.

In this study, multiple machine learning models were constructed to predict hemorrhagic transformation after AIS based on the radiomics features of the infarcted area in CT plain scan images. Then the optimal machine learning model and clinical factors were then combined to construct a nomogram for the risk assessment of hemorrhagic transformation in individualized AIS patients.

Patients

The study was conducted following the Declaration of Helsinki and was approved by the ethics committee of participating hospitals, with the requirement for informed consent waived.

This is a two-center study that retrospectively collected patients with AIS from January 2015 to August 2022. The diagnostic criteria for AIS and hemorrhagic transformation are based on the criteria established by the European Cooperative Acute Stroke Study (ECASS) [15]. The inclusion criteria were as follows: (a) the time of onset was more than 1 day and less than 14 days; (b) responsible lesions could be seen in cranial CT; (c) patients did not receive recanalization therapy such as intravenous thrombolysis or interventional thrombectomy; and (d) reexamination of CT in 7 days. The exclusion criteria were as follows: (a) there is a history of cerebral hemorrhage, cerebrovascular malformation, brain tumor, brain trauma, or brain surgery; (b) participating in other studies and receiving additional drugs or treatment; (c) image artifacts or other factors affect image quality; and (d) AIS has occurred for more than 14 days. According to the stratified sampling, the enrolled datasets were divided into training group and validation cohort at 7:3. The flow chart of the research design is displayed in Fig. 1

Clinical characteristics

The following clinical characteristics were collected for patients with AIS: age, gender, previous vascular risk factors (such as hypertension, diabetes, hyperlipidemia, smoking, etc.), atrial fibrillation, past stroke, transient ischemic attack (TIA), blood pressure, blood glucose, blood lipids, previously used drugs, National Institute of Health stroke scale (NIHSS) score at admission, and head CT plain scan before admission. Improve the relevant examinations such as carotid ultrasound, according to the results of these examinations to classify Trial of Org 10172 in Acute Stroke Treatment (TOAST) stroke subtypes. Head CT plain scan was performed on the 7th day after treatment, and CT was re-examined at any time when the symptoms of stroke worsened. According to the re-examination of head CT, the patients were divided into hemorrhagic transformation group and non-hemorrhagic transformation group.

CT data acquisition

All the head CT plain scan images of were obtained from 6 multislice CT scanners from Department of CT/MRI, The Second Affiliated Hospital of Fujian Medical University. (Brilliance 64 or Brilliance iCT, Incisive 60, Philips Healthcare; SOMATOM Definition DS CT, Siemens Healthcare; Optima CT520 or Optima CT620, GE Healthcare). The scanning parameters were as follows: tube voltage: 120 kVp; tube current: Optima CT520 226 mA, Brilliance 64 228 mA, Incisive 60 230 mA, automatic tube current modulation was using by other CT scanner; Feld of view (FOV): FOV was adapted to the size of patient; matrix size: 512×512; section thickness: 5 mm; section interval: 5 mmNAC regimen.

Cerebral infarction region segmentation

ITK-SNAP (version 3.8.0) was using for manually segmentation of the region of interest (ROI) [16]. The segmentation was carried out along the edge of the infarcted area on each slicer of the CT image（Fig.2）. Reader 1, who was blinded to the patient information, segmented All ROIs. Then, all ROIs were segmented again by the same radiologist after one month for evaluating the intra-observer agreement. Another neurologist segmented all cases over the same period for evaluating the inter-observer agreement. The intra-class correlation coefficient (ICC) was applied to assess the reproducibility of radiomic feature extraction, and two-way mixed-effects model was used to calculate ICC [17]. In this study, radiomic features of ICC > 0.90 were selected and considered a mark of satisfactory intra- and inter-observer agreement.

Radiomic Feature Extraction

A total of 1316 radiomic features were extracted from cerebral infarction region using AK application belonging to PHIgo workstation (Version 1.5.0 R), which can be roughly divided into the following categories: first order statistics, shape-based, gray level cooccurrence matrix (GLCM), gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM), neighboring gray tone difference matrix (NGTDM), and gray level dependence matrix (GLDM).

Classifier Modeling

Max-Relevance and Min-Redundancy (mRMR) and the least absolute shrinkage and selection operator (LASSO) were successively used for radiomic feature selection. Based on the heuristic scoring criteria, radiomic features with ICCs of > 0.90 were ranked by the mRMR method according to their relevance-redundancy index, and the top 10 features with high relevance were retained. The optimized subset of features was then select from the retained features by the LASSO classifier. The radiomics signature was then constructed by multiple logistic regression based on the selected radiomic features. Rad-score was calculated by summing the selected features weighted by their coefficients.

A clinicoradiomics nomogram was generated by combining clinical data and radiomics signature using multivariate logistic regression analysis. A 10-fold cross-validation was performed on the training group to determine the optimal parameter configuration. A clinical model was generated in this process. The Hosmer–Lemeshow test were sued to determine the calibration of models.

Statistical Analysis

The clinical features between hemorrhagic transformation group and non-hemorrhagic transformation group were evaluated using the two independent sample t-test or Mann-Whitney U test accordingly. The predictive performance of machine learning models was evaluated by the receiver operating characteristic (ROC) curve. Differences in area under the ROC curves (AUCs) between the models were evaluated using Delong’s test. In addition, only the machine learning model with the highest AUC of each sequence was used for Delong’s test. P-values of < 0.05 were considered significant. The data was analyzed using SPSS 25 (version 25), R software (version 4.0.4), and MATLAB (version 2022a).

Demographics

A total of 140 patients with AIS were selected for this study. All patients were assigned randomly for the training group (n = 99) or the validation group (n = 41). Of the patients with AIS, 81 were of hemorrhagic transformation and 59 were of stable. Of the 140 patients, 86 were males and 54 were females. The mean age of the patients was 65.2 ± 12.2 years, range 39–93 years.

Radiomic signature construction

Of the 1116 radiomic features, 923 features with intra- and inter-observer ICCs > 0.80 were retained. mRMR eliminates redundant and irrelevant features for stable radiomic features, and 20 features were retained. Based on training group, 10 most significant features were selected to build the radiomic signature by LASSO. 10 most significant feature names and corresponding coefficients are displayed in Fig. 3.

Diagnostic ability of radiomic signature

In the training group, the AUC value of the radiomic signature was 0.86 (95% confidence interval (CI) 0.79–0.93). In the validation group, the AUC value of the radiomic signature was 0.90 (95% CI 0.80–1.00). Further information can be found in Table 1; Fig. 4.

Table 1

The diagnostic performance of radiomic signature in training and validation cohorts
Models	groups	AUC	95% CI	Accuracy	Sensitivity	Specificity
Radiomics signature	Training group	0.86	0.79–0.93	0.79	0.79	0.79
	Validation group	0.90	0.80–1.00	0.83	0.88	0.79
clinicoradiomics nomogram	Training group	0.86	0.78–0.93	0.81	0.71	0.81
	Validation group	0.90	0.80-1.00	0.88	0.80	0.94
Clinical model	Training group	0.64	0.56–0.72	0.69	0.36	0.79
	Validation group	0.63	0.50–0.76	0.68	0.35	0.75
AUC area under the curve, CI confidence interval.

The diagnostic ability of nomogram

Clinical data, including age, sex, hypertension, diabetes, high cholesterol, smoking, atrial fibrillation, stroke history, transient ischemic attack, blood pressure on admission, FBG, serum cholesterol, previous medications and NIHSS, were used for the clinical model, and the clinical data and radiomics signature were used for the clinicoradiomics nomogram. Multivariable analysis revealed that the rad-score and atrial fibrillation were significant independent factors of AIS with hemorrhage transformation (Fig. 5).

For the clinical model, the AUCs with the training group and the validation group were 0.64 (95% CI 0.56–0.72) and 0.63 (95% CI 0.50–0.76), respectively. For the clinicoradiomics nomogram, the AUCs with the training group and the validation group were 0.86 (95% CI 0.78–0.927) and 0.90 (95% CI 0.80-1.00), respectively (Table 1; Fig. 4). The calibration curve was indicative of a strong level of agreement with the actual and predicted diagnosis in both cohorts (Fig. 6). The results of Hosmer-Lemeshow test were not statistically significant, with a p value of 0.33 for the training group and 0.15 for the validation group.

Decision curve analysis (DCA) curves showed that the radiomics signature and clinicoradiomics nomogram were significantly better than the clinical model in predicting AIS with hemorrhage transformation (Fig. 7).

In recent years, imaging has played an important role in the diagnosis, treatment and prognosis of cerebrovascular diseases. Hemorrhagic transformation is closely related to the poor prognosis of patients with AIS, and it is also an important reference index for clinical treatment [18]. Previous studies have established models based on imaging features to predict the risk of hemorrhagic transformation in patients with AIS after thrombolysis or interventional thrombectomy, and showed good predictive efficiency [19–21]. However, as far as we know, there are few studies on predicting the risk of hemorrhagic transformation in patients who do not receive recanalization therapy such as thrombolysis or interventional thrombectomy (accounting for the majority of AIS) based on imaging characteristics.

Similar to previous studies, we also found that atrial fibrillation was an independent risk factor for hemorrhagic transformation [22, 23]. As we all know, the cardiogenic emboli formed by atrial fibrillation are usually large, and it is easy to block the internal carotid artery, middle cerebral artery and other large arteries to cause large area cerebral infarction. A large number of literatures show that large area cerebral infarction is more likely to occur hemorrhagic transformation. The mechanism may be that when a large cerebral infarction occurs, brain tissue ischemia and hypoxia lead to severe cytotoxic edema, which exacerbates the disruption of the blood-brain barrier, thus leading to the occurrence of hemorrhagic transformation [19, 24]. However, the results of this study showed that the radiomics signature was less disturbed by clinical factors, and the predictive efficiency was significantly better than the traditional clinical model.

In this study, the clinical model, radiomics signature and clinicoradiomics nomogram were constructed to predict hemorrhagic transformation in AIS patients who failed to receive thrombolysis or interventional thrombectomy. The results showed that the head CT-based radiomics signature and clinicoradiomics nomogram both had good predictive performance, and the validation accuracy was 0.83 and 0.88 respectively. However, according to the DCA curve, the clinical application value of the clinicoradiomics nomogram is consistent with that of the radiomics signature. In the absence of detailed clinical data and hematological indicators at the initial stage of the disease, the risk of hemorrhagic transformation can be predicted only according to the radiomic features based on CT images, which provides an important reference for clinicians to early evaluate the risk of hemorrhagic transformation and make a preliminary treatment plan. Early and accurate identification of high-risk AIS patients with hemorrhagic transformation can enable clinicians to carry out more stringent clinical monitoring and head CT reexamination in the early stage, to prevent hemorrhagic transformation and improve the prognosis of patients [8]. In addition, according to the predicted results of the model, the risk of hemorrhagic transformation in patients with AIS can be stratified to help clinicians make individual and accurate clinical treatment plans.

A limitation of this study is the small number of patients. Studies including larger patient cohorts should be conducted to further demonstrate the results of this study. The development of automatic segmentation of pelvic bone marrow is urgently needed. This study only focused on AIS patients with hemorrhagic transformation, which result in selection bias and lower power.

In short, early prediction of hemorrhagic transformation is very important, and the radiomics signature which was constructed without clinical characteristics can independently predict the hemorrhagic transformation of AIS well. This radiomics signature may change the clinical management plan of patients, improve clinical treatment decisions and benefit more patients.

Ethics approval and consent to participate

The studies involving human participants were reviewed and approved by Second Affiliated Hospital of Fujian Medical University.

Consent for publication

The submitted manuscript is an original and unpublished work of authors and is not under simultaneous consideration for publication by any another journal.

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Competing interests

The authors of this manuscript declare no relationships with any companies, whose products or services may be related to the subject matter of the article.

Funding

This study has received funding by Quanzhou City Science and Technology Program of China (No. 2019N084S).

Authors' contributions

Yin-Hui Huang, Yuan-Zhe Li, and Yi Wang: study design. Ya-Fang Chen and Chi Cai: data interpretation. Yin-Hui Huang and Ya-Fang Chen: manuscript production. You-Yu Lin, Zhi-Qiang Lin, Chun-Nuan Chen, and Mei-Li Yang take responsibility for the integrity of the data analysis. All authors contributed to the article and approved the submitted version.

Acknowledgements

This work was supported by the Nursery Fund Project of the Quanzhou City Science and Technology Program of China (No. 2019N084S). The authors received the technical assistance provided by the Deep Red AI platform.

Collaborators GBDS (2019) Global, regional, and national burden of stroke, 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurol 18:439–458
Ande SR, Grynspan J, Aviv RI, Shankar JJS (2022) Imaging for Predicting Hemorrhagic Transformation of Acute Ischemic Stroke-A Narrative Review. Can Assoc Radiol J 73:194–202
Paciaroni M, Bandini F, Agnelli G et al (2018) Hemorrhagic Transformation in Patients With Acute Ischemic Stroke and Atrial Fibrillation: Time to Initiation of Oral Anticoagulant Therapy and Outcomes. J Am Heart Assoc 7:e010133
van Kranendonk KR, Treurniet KM, Boers AMM et al (2019) Hemorrhagic transformation is associated with poor functional outcome in patients with acute ischemic stroke due to a large vessel occlusion. J Neurointerv Surg 11:464–468
Hong JM, Kim DS, Kim M (2021) Hemorrhagic Transformation After Ischemic Stroke: Mechanisms and Management. Front Neurol 12:703258
Demaerschalk BM, Kleindorfer DO, Adeoye OM et al (2016) Scientific Rationale for the Inclusion and Exclusion Criteria for Intravenous Alteplase in Acute Ischemic Stroke: A Statement for Healthcare Professionals From the American Heart Association/American Stroke Association. Stroke 47:581–641
de Campos AM, Carvalho A, Rodrigues M et al (2020) Ultra-early improvement after endovascular thrombectomy and long-term outcome in anterior circulation acute ischemic stroke. J Neurol Sci 412:116665
van Kranendonk KR, Treurniet KM, Boers AMM et al (2019) Clinical and Imaging Markers Associated With Hemorrhagic Transformation in Patients With Acute Ischemic Stroke. Stroke 50:2037–2043
Kalinin MN, Khasanova DR, Ibatullin MM (2017) The hemorrhagic transformation index score: a prediction tool in middle cerebral artery ischemic stroke. BMC Neurol 17:177
Krishnamoorthy S, Singh G, Jose KJ et al (2022) Biomarkers in the Prediction of Hemorrhagic Transformation in Acute Stroke: A Systematic Review and Meta-Analysis. Cerebrovasc Dis 51:235–247
Yu Y, Xie Y, Thamm T et al (2020) Use of Deep Learning to Predict Final Ischemic Stroke Lesions From Initial Magnetic Resonance Imaging. JAMA Netw Open 3:e200772
Nishi H, Oishi N, Ishii A et al (2019) Predicting Clinical Outcomes of Large Vessel Occlusion Before Mechanical Thrombectomy Using Machine Learning. Stroke 50:2379–2388
Álvarez-Sabín J, Maisterra O, Santamarina E, Kase CS (2013) Factors influencing haemorrhagic transformation in ischaemic stroke. Lancet Neurol 12:689–705
Jensen M, Schlemm E, Cheng B et al (2020) Clinical Characteristics and Outcome of Patients With Hemorrhagic Transformation After Intravenous Thrombolysis in the WAKE-UP Trial. Front Neurol 11:957
Hacke W, Kaste M, Fieschi C et al (1998) Randomised double-blind placebo-controlled trial of thrombolytic therapy with intravenous alteplase in acute ischaemic stroke (ECASS II). Second European-Australasian Acute Stroke Study Investigators. Lancet 352:1245–1251
Yushkevich PA, Piven J, Hazlett HC et al (2006) User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 31:1116–1128
Koo TK, Li MY (2016) A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med 15:155–163
Xie G, Li T, Ren Y et al (2022) Radiomics-based infarct features on CT predict hemorrhagic transformation in patients with acute ischemic stroke. Front Neurosci 16:1002717
Zhai H, Liu Z, Wu S, Cao Z, Xu Y, Lv Y (2022) Predictive value of magnetic resonance imaging-based texture analysis for hemorrhage transformation in large cerebral infarction. Front Neurosci 16:923708
Jiang L, Zhou L, Yong W et al (2021) A deep learning-based model for prediction of hemorrhagic transformation after stroke. Brain Pathol e13023
Liu J, Tao W, Wang Z, Chen X, Wu B, Liu M (2021) Radiomics-based prediction of hemorrhage expansion among patients with thrombolysis/thrombectomy related-hemorrhagic transformation using machine learning. Ther Adv Neurol Disord 14:17562864211060029
Wen L, Zhang S, Wan K, Zhang H, Zhang X (2020) Risk factors of haemorrhagic transformation for acute ischaemic stroke in Chinese patients receiving intravenous thrombolysis: A meta-analysis. Medicine (Baltimore) 99:e18995
Ge WQ, Chen J, Pan H, Chen F, Zhou CY (2018) Analysis of Risk Factors Increased Hemorrhagic Transformation after Acute Ischemic Stroke. J Stroke Cerebrovasc Dis 27:3587–3590
Yoshimura S, Sakai N, Yamagami H et al (2022) Endovascular Therapy for Acute Stroke with a Large Ischemic Region. N Engl J Med 386:1303–1313

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

The value of CT-based radiomics in predicting hemorrhagic transformation in acute ischemic stroke patients without recanalization therapy

Status:

Version 1

Abstract

Figures

Background

Methods

Patients

Clinical characteristics

CT data acquisition

Cerebral infarction region segmentation

Radiomic Feature Extraction

Classifier Modeling

Statistical Analysis

Results

Demographics

Radiomic signature construction

Diagnostic ability of radiomic signature

The diagnostic ability of nomogram

Discussion

Declarations

References

Additional Declarations

Status:

Version 1