Article Text
PDF
Abstract
Background Thrombus composition has been shown to be a major determinant of recanalization success and occurrence of complications in mechanical thrombectomy. The most important parameters of thrombus behavior during interventional procedures are relative fractions of fibrin and red blood cells (RBCs). We hypothesized that quantitative information from admission non-contrast CT (NCCT) and CT angiography (CTA) can be used for machine learning based prediction of thrombus composition.
Methods The analysis included 112 patients with occlusion of the carotid-T or middle cerebral artery who underwent thrombectomy. Thrombi samples were histologically analyzed and fractions of fibrin and RBCs were determined. Thrombi were semi-automatically delineated in CTA scans and NCCT scans were registered to the same space. Two regions of interest (ROIs) were defined for each thrombus: small-diameter ROIs capture vessel walls and thrombi, large-diameter ROIs reflect peri-vascular tissue responses. 4844 quantitative image markers were extracted and evaluated for their ability to predict thrombus composition using random forest algorithms in a nested fivefold cross validation.
Results Test set receiver operating characteristic area under the curve was 0.83 (95% CI 0.80 to 0.87) for differentiating RBC-rich thrombi and 0.84 (95% CI 0.80 to 0.87) for differentiating fibrin-rich thrombi. At maximum Youden-Index, RBC-rich thrombi were identified at 77% sensitivity and 74% specificity; for fibrin-rich thrombi the classifier reached 81% sensitivity at 73% specificity.
Conclusions Machine learning based analysis of admission imaging allows for prediction of clot composition. Perspectively, such an approach could allow selection of clot-specific devices and retrieval procedures for personalized thrombectomy strategies.
- stroke
- thrombectomy
- CT
- embolic
Data availability statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Statistics from Altmetric.com
Introduction
Mechanical thrombectomy has become frequently used after randomized trials have confirmed its efficacy and benefit for neurological outcome in patients with ischemic stroke. We and others have reported three main thrombus components: fibrin, red blood cells (RBCs), and white blood cells (WBCs), with fibrin and erythrocytes mainly determining physical thrombus behavior during interventional procedures.1–7
Moreover, we observed that histological thrombus features contain important information about stroke etiology and can predict interventional and clinical outcome.3 4 7
Manual assessment of thrombus features using admission imaging has been reported to achieve promising results for prediction of thrombus histology.3 8 However, the apparent disadvantages are that such approach is rater dependent and usually considers only one or a few of the many thrombus imaging characteristics; that is, its density in non-contrast CT (NCCT)3 or perviousness measures in CT angiography (CTA).8
We therefore developed a machine learning based classifier that uses several high-end imaging features derived from readily available admission NCCT and CTA. We hypothesized that this approach will allow for prediction of relevant thrombus histology characteristics and might support optimized device selection and therapeutic strategies for individual patients.
Methods
Patients
We retrospectively evaluated data of a prospectively collected cohort of 198 consecutive patients with ischemic stroke who had occlusion in the carotid-T or middle cerebral artery and in whom pretherapeutic CT imaging and thrombectomy with stent retriever was performed. All patients were scanned and treated in the same comprehensive stroke center between December 2013 and February 2016 using the same CT scanner; all patients were ‘mothership’ patients. Retrieved thrombus material was analyzed histologically. Only patients with complete diagnostic and histological workup were included for further analysis; patients with insufficient imaging quality for the required analysis were excluded. In total, 112 patients were included in the study cohort (table 1).
Baseline characteristics of study population. Data are median (IQR) unless indicated otherwise
We primarily investigated the relationship between thrombus histology expressed as a percentage of the main components (fibrin, RBCs) and quantitative image features. In addition, baseline information including sex, age and interventional parameters was obtained from patients’ clinical records and imaging on admission (table 1).
For evaluating the potential of computer vision based classification of thrombi, all samples were divided into subgroups as follows: based on the ratio of fibrin to RBC, a fibrin-rich subsample was defined as the upper third quantile and the RBC-rich subsample was defined as the lower third quantile of the fibrin/RBC ratio. The study was approved by the Ethics Committee of the University Münster and the Westfalian Chamber of Physicians, Münster, Germany. All study protocols and procedures were conducted in accordance with the Declaration of Helsinki.
CT imaging
CT images at admission were acquired on a 2×128 slice scanner (SOMATOM Definition Flash, Siemens Healthcare GmbH, Erlangen, Germany) with the following imaging parameters: NCCT with 120 kV, 280 mA, less than 5.0 mm slice reconstruction and less than 0.5 mm in-plane; CTA: 100–120 kV, between 260 and 300 mA, 1.0 mm slice reconstruction, 0.5 mm collimation, 0.8 pitch, H20f soft kernel, 60 mL highly iodinated contrast medium and 30 mL NaCl flush at 4 mL/s; scan starts 6 s after bolus tracking at the level of the ascending aorta.
Interventional thrombectomy
Interventional thrombectomy was performed by experienced neuroradiologists under general anesthesia. In all patients we used a biplane neuro-X-ray system (Allura Xper FD20/20, Philips, Best, The Netherlands). Interventional therapy included positioning of a 6F guiding catheter in the cervical segment of the internal carotid artery. In all patients a pRESet stent retriever device (Phenox, Bochum, Germany) with a size of 20×4 or 30×6 mm was used. In some cases a combination of proximal and distal aspiration and stent retriever was performed. All angiograms were reviewed postinterventionally and graded using the modified thrombolysis in cerebral infarction score.
Histology and quantification of stainings
Five µm thin sections of paraffin-embedded, formalin-fixed samples were prepared. Sections were stained with hematoxylin and eosin as well as with Elastica van Gieson and Prussian blue following a standard protocol. Using an Olympus BX43 microscope and digital camera, photographs of thrombus material were acquired (magnification 40×). For quantification of erythrocytes, fibrin and other cellular components, ImageJ software (ImageJ 1.47 n, National Institute of Health, Bethesda, MD, USA) was used. Quantification of fibrin, RBCs and WBCs was performed manually (areas covered by the respective cells (%) were measured).
Segmentation of thrombi
Segmentation of thrombi in CTA scans was conducted semi-automatically. Two experienced neuroradiologists (UH and PBS) documented the voxel coordinates of the proximal thrombus ending, the distal thrombus ending and the thrombus middle. Consensus coordinates were calculated as the center point of the coordinates of both readers. The inner-vessel course of the thrombus was approximated using second degree B-spline interpolation of the three points. Regions of interest (ROIs) were generated through spherical dilation of the interpolation line. Two different ROIs were defined: small-diameter ROIs with radius r=2 mm were calculated to capture vessel walls and thrombi; large-diameter ROIs with radius r=4 mm were generated to reflect peri-vascular tissue responses. Final ROIs were visually verified by UH and PBS. To extract ROIs from NCCT images, respective scans were registered to the CTA images using two-step affine algorithms.9 Readers were blinded to all clinical information. UH and PBS have eight years of clinical experience in diagnostic neuroradiology, research with focus on clinical applications of image processing and predictive modelling.
Machine learning based thrombus differentiation
An overview of the proposed machine learning approach for predicting thrombus histological characteristics is given in figure 1; its modules are detailed below.
Conceptual overview of the proposed machine learning approach showing the major processing steps: CT based image acquisition and segmentation (red line: small-diameter thrombus region of interest; green line: large-diameter thrombus region of interest), feature extraction (n=4 x 1218), statistical learning (random forest algorithm) in nested cross-validation and receiver operating characteristic analysis. CTA: CT angiography; NE CT: non-contrast enhanced CT; ROC: receiver operating characteristic; CV: cross validation set.
Feature extraction and importance assessment
Quantitative image features were extracted using the PyRadiomics Python package v2.1.010; proposed default settings were used for the analysis. Extracted features for each ROI and each scan (NCCT and CTA) comprised 14 shape features (based on ROIs), 252 first-order features (18 based on unfiltered images, 144 wavelet decompositions, 90 log-sigma laplacian of Gaussian filtered images) and 966 texture features (82 based on unfiltered images, 544 wavelet decompositions, 340 log-sigma laplacian of gaussian filtered images). In total, 4844 quantitative image features were extracted for each thrombus sample. Image features were sorted according to predictive value based on Gini impurity measures11 for each training dataset in a nested fivefold cross-validation approach.
Machine learning-based thrombus classification
Image-based histological classification of thrombi was evaluated using random forest algorithms (Python scikit-learn environment v0.20.312). Random forest classifiers were shown to have a comparably low tendency to overfit13 and support classification tasks with numerous and heterogeneous predictors. Two classification tasks were evaluated: differentiation of RBC-rich (lower third quantile of fibrin/RBC ratio) versus other thrombi and differentiation of fibrin-rich (upper third quantile of fibrin/RBC ratio) versus other thrombi. Hyperparameters (total number of features, number of trees, maximum depth of the tree, minimum number of samples to split an internal node, number of features considered for splitting (mtry), minimum number of samples at leaf node) were tuned in a nested fivefold cross-validation approach using grid search algorithms. Initial parameters were set to scikit-learn default values. Due to the relatively balanced dataset for both classification tasks with an event rate of 33% (34/112), no additional data augmentation for reducing bias from class imbalance was performed.
Statistical analysis
For categorical data, absolute and relative frequencies are given. Univariable distribution of variables is described by median and IQR. Thrombi with high RBC/fibrin ratio (upper third quantile) versus low RBC/fibrin ratio (lower third quantile) were compared by Mann-Whitney U test for metric outcome variables and by χ2 test for categorical outcome variables.
Receiver operating characteristic (ROC) curves for predicting thrombus histology were derived based on predictions performed on the test sets of the nested fivefold cross validation. Model prediction instability (ie, SD of areas under curves (AUCs)) was evaluated using 10 randomly drawn fivefold cross-validation sets. Confidence intervals for ROC curves were derived using pROC v1.1014 R-package. Generalized classification performance was further assessed at maximum Youden Index cut-off values through sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) (ThresholdROC v2.80 R-package). In addition, the maximum of Matthews correlation coefficient (MCC)15 was calculated. MCC evaluates all fields of the confusion matrix and is considered a favorable metric for unbiased comparisons of binary classifiers,16 with TP=true positive, TN=true negative, FP=false positive and FN=false negative. MCC is defined as:
MCC confidence intervals were computed with the psychometric v2.2 R-package.
Results
Baseline characteristics of study population
Thirty eight of 112 included thrombi had a RBC/fibrin ratio larger than the upper third quantile and 38 of 112 had a RBC/fibrin ratio lower than the lower third quantile. Thus, both subgroups comprised 38/112 (34%) of total samples. The following baseline metrics were significantly different for fibrin-rich and RBC-rich thrombi at p-value less than 0.05: sex (% female) with 58% versus 34% (p-value=0.039); TICI (thrombolysis in cerebral infarction scale17) grade median with 2b (fibrin rich) and 3 (RBC rich) at p-value=0.005; thrombus Hounsfield unit (HU) in NCCT and CTA with 45/80 HU versus 51/61 HU (p-values=0.003/<0.001); RBC % median with 8.5 versus 53 (p-value<0.001) and fibrin % median with 82 versus 40.5 (p-value<0.001).
Machine learning based histology prediction
Hyperparameter tuning
Results of hyperparameter tuning on validation sets were as follows. For fibrin-rich thrombi, grid search algorithms yielded median ROC AUCs of 0.89 (IQR 0.87–0.91); median number of features to use of 75 (IQR 50–100); median maximum depth of trees of 10; median number of features sampled for splitting at each node (mtry) of 1; median minimum number of samples required to be at a leaf node of 1; median minimum number of samples required to split an internal node of 5 (IQR 2–6) and median number of trees of 1000 (IQR 500–1500). For RBC-rich thrombi, median ROC AUCs were 0.86 (IQR 0.84–0.87); median number of features to use of 50 (IQR 25–50); median maximum depth of trees of 10; median number of features sampled for splitting at each node (mtry) of 1 (IQR 1–2); median minimum number of samples required to be at a leaf node of 1; median minimum number of samples required to split an internal node of 2 (IQR 2–5) and median number of trees of 1000 (IQR 500–1500).
Test set model performance
ROC AUC of the test sets was 0.83 (95% CI 0.80 to 0.87) for differentiating RBC-rich thrombi and 0.84 (95% CI 0.80 to 0.87) for differentiating fibrin-rich thrombi (figure 2). At maximum Youden Index cut-off values, RBC-rich thrombi were identified at 77% sensitivity and 74% specificity; for fibrin-rich thrombi the classifier reached 81% sensitivity at 73% specificity (figure 2, table 2).
Receiver operating characteristic curves of machine learning classifiers for differentiation of thrombus histology based on non-contrast CT and CT angiography image data. Curves are derived from results of nested fivefold cross validation (112 samples) using test set results. AUC: area under the curve; RBC: red blood cell.
Classification performance of proposed classifier
Feature importance analyses
Feature importance analyses of the mean top 100 predictors suggest different radiomic signatures of RBC and fibrin rich thrombi (figure 3): whereas identification of RBC-rich thrombi mainly depends on quantitative features extracted from small-diameter ROIs (74%) that directly reflect thrombus/vessel wall characteristics, identification of fibrin-rich thrombi employs small and large diameter ROIs at equal proportions (54% vs 46%). Furthermore, classifiers for differentiation of RBC-rich thrombi use a higher share of CTA-based features (34% vs 19%) and texture features (74% vs 46%). For both classifiers, importance contribution of original unfiltered images is comparatively low versus wavelet and log-sigma filtered images (3% and 6% vs 97% and 94%).
Predictive value of quantitative image features. Pie charts show distribution of ROIs, applied filters and feature classes in top 100 predictors used. CTA: CT angiography; NECT: non-contrast enhanced CT; RBC: red blood cell; ROI: region of interest.
Discussion
Our study shows that a machine learning based classifier using quantitative markers from admission imaging allows for prediction of histological thrombus composition with promising ROC AUC metrics of 0.83 (fibrin rich) to 0.84 (RBC rich). This finding is clinically important as thrombus composition is associated with different interventional and clinical outcomes.2–4 6 7 18 19
Clots with higher fibrin content have a lower respond rate to intravenous lysis, while RBC-rich clots have better responsiveness to mechanical thrombectomy in terms of time to recanalization, number of passes and recanalization success (TICI).20–24 Furthermore it has been described that fibrin-rich thrombi have a higher static coefficient of friction6 with an increased adherence to vessel walls and decreased compressibility. For RBC-rich thrombi, lower friction coefficients were observed, suggesting that erythrocyte-rich thrombi have a higher probability for migration even though this may also be influenced by other factors such as blood pressure, hematocrit, and vessel anatomy.6 However, studies have shown that fibrin-rich thrombi tend to cause higher rates of secondary embolisms,3 so that in those cases it may be even more relevant to use distal embolism protection than in cases with RBC-rich thrombi. Respective results indicate that recanalization success and patient outcome might benefit from a thrombus-specific device selection and interventional strategy.
The proposed computer vision based classification of thrombus histology allows for a preoperative adaptation of interventional devices, which could become more and more important as technical development moves on. We believe that an individually tailored interventional approach with thrombus composition derived from admission imaging will be key for further improving recanalization rates, as well as rates of secondary embolisms from thrombus fragmentation3 and in the end further improving clinical outcomes.
In line with our results, prior studies have established a correlation between clot histology and clot image features in NCCT25–27 : it has been shown that RBC-rich clots tend to have higher attenuation due to higher hemoglobin concentration within the clot. Furthermore, the absence of a hyperdense middle cerebral artery sign may be an indicator for fibrin-domiant clots.5 28 29
To date, computer vision based classification of histology has been reported for in vitro thrombus analogues using NCCT, dual-energy CT and MRI imaging.30–34 However, a proof of concept using clinical images of patients undergoing mechanical thrombectomy has not been documented yet.
Recently, thrombus perviousness was found to be associated with higher RBC density and lower fibrin amounts, and to predict clinical and interventional outcome.27 35 Another study found that stroke cause may be associated with thrombus perviousness.36 However, the assessment of thrombus perviousness in clinical routine is complicated and so recently a CTA index was introduced as a simplified assessment method to predict angiographic and clinical outcome.37 Also our results suggest a high predictive value of CTA-based image features and hence confirm these findings.
Our study has limitations partly attributable to its single center retrospective design. Even though this is a large cohort of histologically analyzed human thrombi, the total number is still small and so our results should be confirmed by multicenter prospective studies. However, low variability of results across different validation sets suggests sufficient robustness for assessing general feasibility and limitations of machine learning based histology classification. Furthermore, the analysis has general limitations typically associated with quantitative radiomics based classification: differences in image acquisition settings (eg, size of the field of view, gantry tilt, contrast agent triggering), underfitting or overfitting of machine learning algorithms and ground truth misclassifications. These limitations could distort classification and may reduce generalizability of results. Bias of these factors was minimized through use of NCCT and CTA scans acquired by the same scanner, the application of random forest algorithms that are comparably stable with regards to overfitting, and use of established methods for quantification of histological characteristics. The risk of overfitting was further reduced by employing a previously described iterative nested fivefold cross-validation approach. Further study-specific limitations were as follows. First, the manual definition of thrombus coordinate areas still implies a certain degree of observer dependence within the machine learning process. To minimize its influence, we derived consensus coordinates from two experienced neuroradiologists. Further, it was shown that radiomic features are comparably stable with regards to variations in segmentations.38 39 Second, thresholds for class assignments were set at an arbitrary split of 1/3 to 2/3. However, this arbitrary split allows a valid feasibility assessment of our hypothesis through relatively balanced class occurrences, avoidance of threshold optimization, and a most generalized proof of functionality. Third, in some cases a combination of proximal and distal aspiration and stent retriever thrombectomy was performed. The device type as well as the combination of different techniques may influence the type of the retrieved clots and may therefore alter clot histology ex post. Fourth, all images were acquired using the same scanner model. Hence, specific image acquisition and reconstruction characteristics of the employed scanner may reduce generalizability of results. However, CT imaging parameters are calibrated to specific Hounsfield units and inter-scanner comparability of images can be assumed.
Conclusion
Machine learning based analysis of admission imaging allows for accurate prediction of clot composition. In clinical routine, this could allow selection of clot-specific devices and adaption of thrombectomy strategies to improve recanalization rates and times, reduce the number of passes and adverse events, and thereby optimize patient outcome.
Data availability statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Ethics statements
Patient consent for publication
References
Footnotes
Twitter @Fie0815
UH and PBS contributed equally.
Correction notice This article has been corrected since it first published. The provenance and peer review statement has been included.
Contributors Substantial contributions to conception and design: UH, HK, PBS, JF, MW, MNP. Acquisition and analysis and interpretation of data: AJ, JM, MG, KS, JN, GB, LM, AB, UH. Drafting a significant portion of the manuscript or figures: HK, UH, PBS.
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests None declared.
Provenance and peer review Not commissioned; externally peer reviewed.