Computed Tomography Radiomic Signatures Associated With Neutrophil Extracellular Trap Enrichment and First‐Pass Outcome in Ischemic Stroke Thrombi

Abstract

Background

Radiomic and transcriptomic analyses have independently identified features linked to mechanical thrombectomy (MT) outcomes. Here, we integrated paired radiomics/transcriptomics of stroke clots to identify neutrophil extracellular trap (NET) enrichment as a predictor of first‐pass MT success, assessing the potential to noninvasively detect NET enrichment using prethrombectomy computed tomography imaging.

Methods

We performed radiomic/transcriptomic analyses of 32 stroke clots retrieved by MT. Clots were segmented from pre‐MT computed tomography angiography and noncontrast computed tomography scans, and radiomic features (RFs) were extracted using PyRadiomics. Differentially expressed genes were identified between modified first‐pass effect (mFPE) success and failure using the criteria of log(fold‐change) ≥1.5 and q <0.05. RFs significantly different between mFPE outcomes were identified. A NET enrichment score was computed from expression data, and RFs that differed significantly between low‐ and high‐NET‐enriched clots were selected to construct an RF signature predictive of NET enrichment. Immunofluorescence was completed on clots to provide ground truth NET labeling.

Results

A total of 44 differentially expressed genes were identified between mFPE outcomes. NET formation, neutrophil degranulation, and the NET signaling pathway were among the most enriched gene ontologies in the mFPE failure group, with related genes downregulated in the mFPE success group. Forty RFs were significantly different between mFPE outcomes. Of these, 4 were found to be predictive of clot NET enrichment. Immunofluorescence validated that transcriptomic NET signatures accurately reflected NET presence within clot tissues.

Conclusions

NET enrichment in clots is associated with reduced mFPE success. With further validation, RFs extracted from prethrombectomy computed tomography imaging may serve as noninvasive biomarkers of clot NET content to aid in preprocedural MT decision‐making.

Nonstandard Abbreviations and Acronyms

AIS

acute ischemic stroke

DEG

differentially expressed gene

FOS

first order statistics

FP

fibrin‐platelet

KEGG

Kyoto Encyclopedia of Genes and Genomes

mFPE

modified first‐pass effect

MT

mechanical thrombectomy

nCCT

noncontrast computed tomography

NET

neutrophil extracellular trap

NGTDM

neighboring gray tone difference matrix

RF

radiomic features

Research Perspective.

What Is New?

• This study introduced a paired radiomic–transcriptomic analysis of stroke clots that identified neutrophil extracellular trap enrichment as a novel biologic correlate of first‐pass effect during thrombectomy.

• It demonstrated, for the first time, that specific computed tomography‐derived radiomic features could potentially be used as noninvasive predictors of neutrophil extracellular trap enrichment within stroke clots.

What Question Should Be Addressed Next?

• Future research should validate these radiomic neutrophil extracellular trap signatures in larger, multicenter cohorts and determine whether such imaging biomarkers can guide individualized thrombectomy strategies.

Acute ischemic stroke (AIS) poses a significant global health burden, necessitating rapid and effective interventions to mitigate significant morbidity and mortality. ¹ Mechanical thrombectomy (MT) has emerged as a pivotal therapeutic approach, offering a lifeline for patients afflicted by large vessel occlusions. ² The success of MT, particularly on the first pass, represents a critical milestone in achieving optimal outcomes. ³ To best predict likelihood of MT success, studies have focused on learning what properties of a clot make it challenging to remove mechanically. A large body of work has been dedicated to investigating whether clot radiomics from prethrombectomy computed tomography (CT) or clot composition (percentage of red blood cell, white blood cell [WBC], and fibrin‐platelet [FP] meshwork) determined from digital histology can predict MT first‐pass outcome. ⁴ , ⁵ , ⁶ , ⁷ Although some findings have had consensus, such as stiffer clots being less amenable to retrieval, ⁸ , ⁹ reports on which biological components of the clot are important are inconsistent. ¹⁰ , ¹¹ It has been suggested that percentage composition, alone, is insufficient to decipher the biology of AIS clots because composition overlooks important factors such as the density of FP meshwork, the localization of WBCs along the red blood cell ‐FP interface, and the enrichment of prothrombotic phenomenon such as neutrophil extracellular traps (NETs). ¹² , ¹³

NETs are networks of extracellular DNA, antimicrobial enzymes, and histone that serve a vital antimicrobial function in the body. ¹⁴ However, ischemic conditions such as those in stroke can facilitate rapid NET formation wherein NETs create a hypercoagulable state and reinforce thrombus microstructure. In the in vitro clot setting, NETs have been observed to bind fibrin, increase crosslinking, and produce denser microstructure. ¹⁵ The compaction that results increases clot stiffness and resistance to both mechanical disruption (MT) and dissolution (tissue plasminogen activator [tPA] resistance). ¹⁶ Additionally, NET DNA strands with a dense negative charge serve as a scaffold for other blood components (FP, red blood cell) whereas enzymes elicit further immune response. In vitro studies have also shown that the addition of DNase I to tPA can help resolve clots into the blood flow, supporting the concept that the NET DNA backbone is a key component altering clot microstructure.

In vitro studies have demonstrated that NETs significantly alter the biophysical properties of thrombi. ¹⁷ , ¹⁸ In addition, ongoing clinical trials are investigating whether DNA‐targeting therapeutics can enhance early reperfusion and improve outcomes in patients with large vessel occlusion stroke. These agents have demonstrated efficacy in degrading NETs and improving outcomes in conditions such as acute respiratory distress syndrome and hypercoagulable states. ¹⁹ Given this, quantifying NET enrichment in AIS thrombi may help identify patients who would benefit most from adjunctive therapies (eg, tPA combined with DNase) or guide the selection of MT devices. Currently, NET detection in stroke clots relies solely on postretrieval immunofluorescence staining, and no validated imaging‐based biomarkers exist for preprocedural assessment. To address this gap, we analyzed NET enrichment in AIS thrombi and investigated its association with radiomic features, representative of clot organization, derived from prethrombectomy CT imaging. If validated in larger cohorts, a radiomic signature of NET content could enable personalized treatment strategies within the critical stroke intervention window.

METHODS

Patient Data

All data, methods, and study information for this work are available and are included in this publication. This study was approved by our institute’s institutional review board (study 00002092). All methods were performed in accordance with the approved protocol, and written informed consent was obtained for all subjects. Clot samples and imaging were collected as previously described from patients receiving MT by either stent retriever, aspiration, or a combined therapy between November 2018 and November 2020 at our institution. ¹³ Patients considered for this study met the following criteria: (1) They had undergone MT for AIS, (2) had pretreatment CT available with sufficient image quality and an identifiable clot, (3) had a retrieved clot of sufficient size and quality for microCT and histology, and (4) did not receive intra‐arterial tPA. MT first‐pass outcomes were defined using the modified first‐pass effect (mFPE) definition of thrombectomy success, which in the context of endovascular thrombectomy is defined as complete or near‐complete reperfusion (Thrombolysis in Cerebral Infarction score of 2b/2c/3) on the first attempt with a thrombectomy device. ³

Clot mRNA Extraction and Differential Expression Analysis

Clot transcriptomes were obtained from those analyzed in our previous work that established a pipeline for obtaining useful RNA from AIS blood clots. ²⁰ In brief, clot samples retrieved by MT and stored in RNALater were subjected to RNA extraction using the Chemagen magnetic bead extraction protocol on a PerkinElmer Chemagic 360. A total of 48 samples were used for RNA sequencing based on sequencing quality metrics. In this study, a subset of n=32 clot transcriptomes for which histology was also available was examined to determine if expression profiles were unique to MT first‐pass outcomes. On average, these clot RNA samples yielded 22 ng/μL after concentration and a 260/280 ratio of 2.1. RNA integrity numbers were not quantifiable on all samples due to initial low concentrations, so RNA quality was assessed by manual inspection of the electrophoresis gels. For those that had sufficient bioanalyzer traces, RNA integrity numbers ranged from 3.3 to 9.5. After sequencing, the output demonstrated an average of 35.5M sequences per sample and an alignment rate of 95.7% on average.

Differential gene expression analysis was completed between patient samples that did or did not achieve mFPE using a Fisher’s exact test in edgeR, as previously described. ²¹ , ²² Significant differentially expressed genes (DEGs) were identified using the following criteria: 50% minimum expression among samples, absolute log(fold‐change) ≥1.5, and q <0.05 after false discovery correction by the Benjamini–‐Hochberg procedure. All gene candidates were plotted as a volcano plot, and DEGs were denoted as colored dots based on whether they were up‐ or downregulated. Using pheatmap in R, DEGs were visualized as a heatmap and hierarchically clustered based on Euclidean distance between observations. ²³

Ontology Analyses and Calculation of Clot NET Enrichment

To understand the biology of mFPE‐associated DEGs, we performed gene set enrichment analyses using ClusterProfiler ²⁴ and QIAGEN ingenuity pathway analysis (QIAGEN Inc., https://digitalinsights.qiagen.com/). For ClusterProfiler, we queried the Kyoto Encyclopedia of Genes and Genomes (KEGG ²⁵ ) and ReactomePA databases using our DEG set. ²⁶ For ingenuity pathway analysis, we specifically analyzed enrichment of the disease and biological function terms (terms that had q‐value<0.05 and 3 or more input genes were considered).

Based on the results of ontology analysis, we further sought to compute the per sample enrichment of NET formation based on clot transcriptomes. To achieve this, we pulled a comprehensive list of all genes defining the KEGG “Neutrophil Extracellular Trap Formation” term (hsa04613) using keggGet. ²⁷ We then computed the Single Sample Gene Set Enrichment Analysis score for each sample using gsva ²⁸ (Gene Set Variation Analysis) and normalized the set of scores between zero and one as recommended. High and low NET enrichment groups were stratified by the median NET score (0.40 as cutoff). All analysis was completed in R v4.4.2.

Immunofluorescence Labeling and Imaging of Neutrophil Extracellular Traps in Retrieved Clots

To provide ground truth values of clot NET enrichment, we completed immunofluorescence staining, whole‐slide imaging, and whole‐slide quantification of MT‐retrieved clot tissues. Clot histological processing and immunofluorescence staining were completed as previously described. ²⁹ In brief, formalin‐fixed paraffin embedded tissues were sectioned at 4 μm thick and mounted on clear glass slides. Indirect immunofluorescence labeling of CitHis (citrullinated histone) was completed using Rabbit Anti Citrullinated Histone (ab281584, 1:200 μg/mL, Abcam, Cambridge, UK), donkey blocking serum, and Donkey Anti Rabbit AF647 secondary antibody (ab150075, 1:200 μg/mL, Abcam, Cambridge, UK). Slides were then DAPI counterstained, and #1.5 coverslips were applied. Whole‐slide fluorescence images were acquired at 20× on a Leica DM6B Fluorescence Microscope (Leica Biosystems, Lincolnshire, IL). The DAPI and Y5 filter cubes were used to image DAPI and AF647 (CitHis), respectively (exposure time was 100–200 ms for DAPI and 400 ms for AF647).

The percentage of CitHis, an established measure of clot NET enrichment, ³⁰ , ³¹ , ³² was then computed as follows. The blue (DNA) and red (CitHis) channels of fluorescence whole‐slide images were extracted, and a top‐hat filter was applied to each to correct for uneven illumination. Global mean‐based thresholds were applied to each channel to segment true‐positive signal from background and produce binary masks for DNA and citrullinated histone. The intersection of DNA and CitHis regions was found by taking the intersection (logical AND) of the respective masks. This region represented the NET‐positive region. The area of CitHis within the NET‐positive area was indexed to the total clot area to yield %CitHis.

CT Image Analysis

For all patients, CT was performed on an Aquilion ONE scanner (Canon Medical Systems). In this study, CT angiography (CTA) and noncontrast CT (nCCT) images with the lowest slice thickness (0.5 mm) and highest resolution were used for analysis. As previously described, 3D Slicer was used to register CTA and nCCT images, manually segment and reconstruct patient vasculatures, and isolate the clot region of interest. ³³ Clot annotation was completed by 2 experienced annotations, and interrater agreement was assessed by Dice Coefficient. PyRadiomics, an open‐source Python package, was used to extract clot morphology and textural features from segmented clot regions in both the CTA and nCCT images (bin size=25, no resampling given uniform slice thickness and scanner). In addition, textural feature ratios were computed by dividing the value of a given radiomic from CTA by its value in nCCT, yielding radiomic feature ratios (CTA:nCCT). Textural features included first order statistics (FOS), gray‐level run length matrix, gray‐level dependence matrix, gray‐level cooccurrence matrix, gray‐level size zone matrix, and neighboring gray tone difference matrix (NGTDM). Formulae for radiomic features (RFs) are available in the PyRadiomics documentation (https://pyradiomics.readthedocs.io/). Combined, CTA, nCCT, and CTA:nCCT (ratio) clot radiomic features totaled 293 features (93 per modality, 93 ratios, and 14 shape). Feature values were normalized between 0 and 1.

Statistical Analysis

All statistical analysis was completed in R v4.4.2 and statistical significance was considered at α<0.05. Statistical analysis was completed to summarize patient demographics and clinical variables, as well as to identify significantly different RFs between first‐pass outcomes. In addition, statistical testing was completed to assess significant difference in percentage CitHis, as well as to identify radiomic features significantly different between low and high NET enriched clots. Normality was assessed using the Anderson Darling Test Statistic and when normally distributed, equal variance was assessed using Bartlett’s test. Based on these results, group comparisons were performed using Student’s t test (parametric, equal variance), Welch’s t test (parametric, unequal variance), or the Mann–Whitney U test (nonparametric).

To evaluate the predictive capability of radiomics for clot NET enrichment, a multivariable logistic regression model with 3‐fold cross‐validation and feature selection was developed. Candidate predictors were CTA, nCCT, CTA:nCCT ratio RFs that differed significantly between low and high NET clots following Benjamini–Hochberg false discovery rate correction (q<0.05). Features were scaled to the (0–1) interval before modeling (see Methods: CT Image Analysis), and the feature subset that maximized model area under the receiver operating characteristic curve (AUC) was selected. For the final model, a receiver operating characteristic curve was generated, and the AUC and corresponding CI were calculated using 1000 bootstrap resamples (boot.n=1000) via the roc() function in the pROC package. The final selected features were visualized in a bar plot ranked by relative importance.

RESULTS

Summary of Patient Population

A summary of characteristics including statistics for the 32 patients (59.3% female) included in this study is provided in Table 1. The mean±SD age of patients was 71.3±14.7 years, and National Institutes of Health Stroke Scale score at presentation was 16.2±7.2. In our cohort, 3.1% of strokes were treated with stent retriever thrombectomy, 65.6% with aspiration, and 31.2% with combination therapy (Solumbra). Occlusions were primarily located on the right (65.6%); 9.4% of clots were located in the internal carotid artery, 25% at the internal carotid artery/M1 junction, 37.5% in the middle cerebral artery M1 region, 18.7% in the middle cerebral artery M2 region, and 9.4% at the basilar. In all, 50% of thrombectomy procedures achieved mFPE. There were no statistically significant differences (all P>0.05) among demographic or clinical variables between the mFPE and no mFPE groups (based on 2‐sample or chi‐square tests).

Table 1.

Baseline Patient Characteristics^*

Demographic data	mFPE (n=16)	No mFPE (n=16)
Sex, female	8 (50.0%)	11 (68.8%)
Age, y	66.3±15.3	76.4±12.1
Past medical history
Body mass index, kg/m2	28.1±4.4	30.7±13.9
Hypertension	12 (75.0%)	14 (87.5%)
Hypercholesterolemia	8 (50.0%)	9 (56.3%)
Heart disease	5 (31.3%)	5 (31.3%)
Deep vein thrombosis	3 (18.8%)	3 (18.8%)
Stroke	2 (12.5%)	3 (18.8%)
Cancer	4 (25.0%)	3 (18.8%)
Diabetes	5 (31.3%)	4 (25.0%)
Arthritis	1 (6.3%)	0 (0.0%)
Asthma	1 (6.3%)	1 (6.3%)
Treatment details
Stent retriever	0 (0.0%)	1 (6.3%)
Aspiration	9 (56.3%)	12 (75.0%)
Aspiration and stent retriever	7 (43.8%)	3 (18.8%)
Stroke presentation
Right‐sided occlusion	10 (62.5%)	11 (68.8%)
Left‐sided occlusion	4 (25.0%)	4 (25.0%)
Basilar occlusion	2 (12.5%)	1 (6.3%)

mFPE indicates modified first‐pass effect.

^*Characteristics are summarized as mean±SD or n (%).

Transcriptomic Signatures Associated With Modified First‐Pass Effect

To identify clot transcriptomic signatures associated with mFPE, we performed differential gene expression analysis comparing patients who achieved mFPE to those who did not. A total of 44 DEGs were identified, with 1 gene upregulated and 43 downregulated in the mFPE success group (Figure 1A and 1B, Table 2). The sole gene with significantly increased expression in successful first‐pass cases was CRIP2 (q=0.031). The top 3 DEGs with significantly decreased expression were MMP8 (q<0.001), MS4A3 (q<0.001), and TCN1 (q=0.004).

Figure 1. Clot gene expression differences between mFPE success and failure thrombectomy attempts.

A, A volcano plot of differentially expressed genes shows 43 genes were downregulated, and 1 gene was upregulated in mFPE successes. B, Hierarchical clustering of differentially expressed genes shows separation between mFPE successes and failures. C, Kyoto Encyclopedia of Genes and Genomes analysis of 44 differentially expressed genes identified pathways enriched in mFPE failures, including neutrophil extracellular trap formation. D, ReactomePA analysis of 44 differentially expressed genes identified pathways enriched in mFPE failures including neutrophil degranulation. GTPase indicates guanosine triphosphate; mFPE, modified first‐pass effect; and NS, not significant.

Table 2.

Differentially Expressed Genes Between mFPE Outcomes

Gene name	Gene ID	LogFC	q‐value
MMP8	ENSG00000118113.11	−4.6737	<0.001
MS4A3	ENSG00000149516.13	−3.11496	<0.001
TCN1	ENSG00000134827.7	−2.92338	0.004
MAP2K6	ENSG00000108984.13	−2.12872	0.006
CRISP3	ENSG00000096006.11	−3.42722	0.006
DEFA3	ENSG00000239839.5	−3.62578	0.008
PF4V1	ENSG00000109272.3	−1.87549	0.011
ANXA3	ENSG00000138772.12	−2.33949	0.011
PI3	ENSG00000124102.4	−2.63576	0.011
CAMP	ENSG00000164047.4	−3.00313	0.011
CEACAM8	ENSG00000124469.10	−3.36298	0.011
LTF	ENSG00000012223.12	−3.3819	0.011
SULT1B1	ENSG00000173597.8	−2.60394	0.012
DEFA4	ENSG00000164821.4	−3.03634	0.015
FGL2	ENSG00000127951.5	−1.62701	0.019
CARD17	ENSG00000255221.2	−2.32638	0.021
RGPD1	ENSG00000187627.14	−3.18508	0.021
CR1	ENSG00000203710.10	−1.82596	0.024
S100A9	ENSG00000163220.10	−1.87246	0.024
COL17A1	ENSG00000065618.16	−3.15841	0.024
S100A8	ENSG00000143546.9	−1.95813	0.025
MME	ENSG00000196549.10	−2.01133	0.025
1‐Mar	ENSG00000186205.12	−2.03505	0.025
BPI	ENSG00000101425.12	−2.37102	0.027
KRT23	ENSG00000108244.16	−1.79118	0.027
PGLYRP1	ENSG00000008438.4	−2.39498	0.027
GPR141	ENSG00000187037.8	−1.62294	0.028
SLPI	ENSG00000124107.5	−1.94273	0.028
PROK2	ENSG00000163421.8	−2.06156	0.028
CLEC12A	ENSG00000172322.13	−2.20877	0.028
CRIP2	ENSG00000182809.10	1.502318	0.031
WLS	ENSG00000116729.13	−1.73543	0.031
MPO	ENSG00000005381.7	−1.85026	0.031
CD24	ENSG00000272398.5	−2.25621	0.031
GALNT14	ENSG00000158089.14	−2.74895	0.031
TRPM6	ENSG00000119121.21	−1.70957	0.038
KCNJ2	ENSG00000123700.4	−1.93854	0.040
ALPL	ENSG00000162551.13	−2.3972	0.040
RNASE3	ENSG00000169397.3	−2.50991	0.040
CTSG	ENSG00000100448.3	−3.45384	0.040
SMARCD3	ENSG00000082014.16	−1.60916	0.043
LILRA5	ENSG00000187116.13	−1.54108	0.044
FAM65B	ENSG00000111913.15	−1.82686	0.047
CNTNAP3	ENSG00000106714.17	−1.62974	0.047

LogFC indicates log fold change; and mFPE, modified first‐pass effect.

Neutrophil Extracellular Trap Formation and Signaling Are Associated With Thrombectomy Failure

To better understand the biological significance of DEGs associated with mFPE, we completed gene set enrichment analysis with KEGG, ReactomePA, and ingenuity pathway analysis databases. From KEGG, the 2 most enriched pathways were Staphylococcus aureus infection (q=0.005) and NET formation (q=0.018) (Figure 1C). ReactomePA also emphasized neutrophil activity, with top terms of neutrophil degranulation (q<0.001) and antimicrobial peptides (q<0.001) (Figure 1D). When the single‐sample gene set enrichment analysis score for KEGG NET formation was plotted for each of the mFPE successes and failures, increased NET formation in mFPE failures was observed (Figure S1A) and confirmed on immunofluorescence (negative control staining, Figure S1B). Gene set enrichment analysis with ingenuity pathway analysis further supported these findings, with the NET signaling pathway identified as the most enriched pathway in mFPE failures (negative value in mFPE successes) (Figure S1C).

Clot Radiomics Are Significant Predictors of mFPE

We next sought to identify CT radiomic signatures associated with mFPE outcomes. Interrater assessment of clot annotations by 2 experienced annotators demonstrated high agreement, with a mean dice coefficient of 0.80±0.09 (Figure 2A). A total of 40 RFs computed from CTA and nCCT clot segmentations were significantly different between mFPE outcomes (Table 3). Of these RFs, 34 were computed from CTA, 1 from nCCT, and 5 from the ratio of CTA to nCCT (Figure 2B). Overall, 3 RFs were common to the CTA and CTA:nCCT ratio analyses (Figure 2B). These RFs were all derived from the gray‐level co‐occurrence matrix and included autocorrelation, joint energy, and maximum probability. Two significant RFs derived from CTA, FOS 90th percentile and gray‐level cooccurrence matrix joint energy, were visualized as heatmaps and superimposed on prethrombectomy CTA for further interpretation. Here, a significant increase in FOS 90th percentile (P=0.021) was observed in mFPE successes, suggesting that clots with a higher 90th Percentile intensity value are more amenable to retrieval (Figure 2C). In comparison, a significant increase in gray‐level cooccurrence matrix joint energy (P=0.017) was observed in mFPE failures, suggesting that clots with greater variation in intensity values, and thus textural complexity, are more likely to result in mFPE failure (Figure 2D).

Figure 2. Clot radiomic differences between mFPE success and failure thrombectomy attempts.

A, Interrater assessment of clot annotations shows high agreement. B, Venn diagram shows significant radiomics common to CTA, nCCT, and ratio of CTA:nCCT feature sets. C, Bar plot shows clot first order statistic 90th percentile measured from CTA as a radiomic significantly different between mFPE success and failure. Radiomic heatmaps visualize the first order statistic 90th percentile in example cases of mFPE success and failure. D, Bar plot shows clot gray‐level co‐occurrence matrix joint energy measured from CTA as a radiomic significantly different between mFPE success and failure. Radiomic heatmaps visualize the gray‐level co‐occurrence joint energy in example cases of mFPE success and failure. 2D indicates 2‐dimensional; CTA, computed tomography angiography; FOS, first order statistic; GLCM, gray‐level co‐occurrence matrix; mFPE, modified first‐pass effect; and nCCT, noncontrast computed tomography.

Table 3.

Radiomics Significantly Different Between mFPE Outcomes

Radiomic feature	P value
FOS 90th percentile, CTA	0.022
FOS energy, CTA	0.033
FOS interquartile range CTA	0.018
FOS kurtosis, ratio	0.019
FOS mean absolute deviation, CTA	0.034
FOS mean, CTA	0.047
FOS robust mean absolute deviation, CTA	0.022
FOS root mean squared, CTA	0.038
FOS skewness, ratio	0.004
FOS total energy, CTA	0.023
FOS uniformity, CTA	0.048
FOS variance, CTA	0.036
GLCM autocorrelation, CTA	0.017
GLCM autocorrelation, ratio	0.031
GLCM cluster prominence, CTA	0.029
GLCM cluster tendency, CTA	0.028
GLCM contrast, CTA	0.049
GLCM difference average, CTA	0.046
GLCM difference entropy, CTA	0.033
GLCM inverse difference, CTA	0.049
GLCM inverse difference moment, CTA	0.047
GLCM inverse difference moment normalized, CTA	0.033
GLCM inverse variance, CTA	0.046
GLCM joint average, CTA	0.019
GLCM joint energy, CTA	0.017
GLCM joint energy, ratio	0.037
GLCM joint entropy, CTA	0.012
GLCM maximum probability, CTA	0.046
GLCM maximum probability, ratio	0.042
GLCM sum average, CTA	0.019
GLCM sum entropy, CTA	0.025
GLCM sum squares, CTA	0.028
GLDM gray‐level variance, CTA	0.038
GLDM high gray‐level emphasis, CTA	0.022
GLRLM gray‐level variance, CTA	0.045
GLRLM high gray‐level run emphasis, CTA	0.023
GLRLM long run high gray‐level emphasis, CTA	0.029
GLRLM short run high gray‐level emphasis, CTA	0.026
Shape maximum 2D diameter row, CTA	0.030
Shape maximum 2D diameter row, noncontrast computed tomography	0.030

2D indicates 2‐dimensional; CTA, computed tomography angiography; FOS, first order statistic; GLCM, gray‐level co‐occurrence matrix; GLDM, gray‐level dependence matrix; GLRLM, gray‐level run length matrix; and mFPE, modified first‐pass effect.

Prethrombectomy CT Radiomics May Predict Clot Neutrophil Extracellular Trap Enrichment

Having identified both clot transcriptomic and radiomic signatures of mFPE outcome, as well as the significance of NET gene set enrichment, we sought to assess whether an association between prethrombectomy clot radiomics and NET enrichment could be identified. We first performed validation of clot NET enrichment by immunofluorescence quantification. After binarizing clot NET scores using the median (a threshold NET score of 0.40), a distribution‐independent measure, we arrived at 14 clots in both the low and high NET enrichment groups (P<0.001) (Figure 3A). We validated this finding through quantification of ground truth immunofluorescence labeling of NET formation in clots retrieved by MT (Figure 3B, Figure S1). For higher magnification images of the immunofluorescence staining of NETs see Figure S1. Here, NET formation was quantified as the percentage of CitHis. When the percentage of CitHis was compared between transcriptome‐defined low and high NET enrichment clots, we identified a significant difference (P=0.007) (Figure 3C).

Figure 3. NET‐enriched clots demonstrate a distinct radiomic signature.

A, Bar plot shows significant difference between mean NET score from transcriptomics for clots found to be low and high NET enriched. B, Immunofluorescence images show whole‐slide scans of clots stained for citrullinated histone, a marker of NET formation, and DNA (DAPI). C, Bar plot shows that clots classified as high NET enrichment based on transcriptomic analysis also demonstrated significantly increased percentage citrullinated histone by whole‐slide image immunofluorescence quantification. D, Receiver operating characteristic curve demonstrates an AUC of 0.859 for a multivariable logistic regression model classifying high and low NET clots based on four selected radiomics. E, Bar plot shows ranking of the 4 selected radiomics based on relative feature importance. F, Radiomic feature maps demonstrate differences in the most and least important selected radiomics between low and high NET enriched clots. Whole‐slide digital pathology for the same clots is provided. Scalebars measure 1 mm. AUC indicates area under the receiver operating characteristic curve; CTA, computed tomography angiography; FOS, first order statistic; GLSZM, gray‐level size zone matrix; LGLE, low gray‐level emphasis; nCCT, noncontrast computed tomography; NET, neutrophil extracellular trap; NGTDM, neighboring gray tone difference matrix; Norm, normalized; and SA, small area.

To demonstrate robustness of the NET score dichotomization, we performed a small sensitivity analysis in which we varied the threshold NET score by ±0.1 and used immunofluorescence to determine if the clots in each group still had significant differences in percentage of CitHis. As shown in Figure S3A through S3C, varying the NET score threshold did not change the significant NET across groups, demonstrating the robustness of the 0.40 threshold.

Next, we identified a preliminary radiomic signature for NET enrichment. Following validation of NET enrichment, we aimed to identify those radiomics significantly different between low and high NET‐enriched clots. Overall, we identified 15 RFs demonstrating significant difference between low and high NET groups (Table 4, q<0.05). The 3 most significant of these radiomics were NGTDM strength on CTA (q=0.014), shape max 2‐dimensional diameter slice (q=0.019), and FOS minimum CTA:nCCT ratio (q=0.023). A logistic regression model based on 4 of these RFs (feature selection) was able to predict NET score with a mean AUC of 0.859, suggesting that prethrombectomy CT radiomics may be able to predict clot NET enrichment (Figure 3D). The 4 radiomics selected by the optimal model were ranked based on their feature importance and included NGTDM strength on CTA, gray‐level size zone matrix small area low gray‐level emphasis CTA:nCCT ratio, FOS mean on nCCT, and NGTDM busyness on nCCT (Figure 3E).

Table 4.

Radiomics Significantly Different Between High and Low NET Enrichment

Radiomic feature	q‐value
NGTDM strength, computed tomography angiography	0.014
Shape maximum 2D diameter slice	0.019
FOS minimum, ratio	0.023
GLSZM small area low gray‐level emphasis, ratio	0.024
Gray‐level run length matrix gray‐level nonuniformity, nCCT	0.033
NGTDM busyness, nCCT	0.033
Gray‐level co‐occurrence matrix cluster shade, ratio	0.035
GLSZM large area emphasis, nCCT	0.038
GLSZM large area low gray‐level emphasis, nCCT	0.040
FOS root mean squared, nCCT	0.044
GLRLM run variance, CTA	0.045
GLSZM size zone nonuniformity normalized, ratio	0.045
NGTDM busyness, CTA	0.048
GLSZM large area high gray‐level emphasis, nCCT	0.049
FOS mean, nCCT	0.049

FOS indicates first order statistic; GLRLM, gray‐level run length matrix; GLSZM, gray‐level size zone matrix; NET, neutrophil extracellular trap; nCCT, noncontrast computed tomography; and NGTDM, neighboring gray tone difference matrix.

When selected features were visualized for low and high NET groups, we observed elevated NGTDM strength on CTA and decreased NGTDM busyness in clots with high NET enrichment (Figure 3F). High NGTDM strength suggests that a clot has regions that are densely packed and homogeneous adjacent to sharply distinct areas, such as contrast interfaces or denser fibrin/erythrocyte cores. Given that this pattern is observed in clots enriched with NETs—which promote compartmentalization of clot components—elevated NGTDM strength likely reflects NET‐dense, fibrin‐rich domains that produce a coarser and more sharply delineated texture. Similarly, NGTDM busyness quantifies the rate of intensity change between neighboring voxels or how rapidly varying the texture is. Thus, low busyness suggests smoother, more uniform regions with gradual or infrequent intensity changes. Overall, increased strength and decreased busyness in high NET clots reflects structural compartmentalization and a more internally uniform clot matrix. This structural pattern found in NET‐rich clots aligns with the pro‐thrombotic NET mechanism, promoting clot stability by facilitating red blood cell aggregation and retention, supporting fibrin linking, and reducing clot permeability and breakdown. ¹⁵ , ³⁴

DISCUSSION

In this study we delved into the molecular underpinnings of AIS clots, with a focus on identifying biological processes associated with successful first‐pass MT. Through transcriptomic profiling of extracted clots, we identified NET formation and the NET signaling pathway as the most significantly enriched processes linked to first‐pass outcome. These findings highlight a critical role for NETs in influencing early thrombectomy outcomes and are consistent with prior studies. ⁶ , ³⁰ , ³² , ³⁴

Our understanding of the role of NETs in AIS clot microstructure, thrombolytic resistance, and thrombectomy failure is evolving. Several prior in vitro studies have demonstrated that NET‐rich clot analogs resist both thrombolysis and mechanical retrieval. ¹⁷ , ¹⁸ Additional studies have shown that combination therapies such as tPA‐DNase I, which target FP and DNA components, improve clot busting capabilities. ³² , ³⁵ Such findings have helped spur clinical trials investigating whether therapeutics with DNase I activity can improve reperfusion rates in large vessel occlusion stroke. In addition to these studies, several published works have demonstrated successful localization and quantification of NETs in retrieved thrombi. In a work by Ducroux et al., ³² it was shown that NETs are constitutively present in AIS thrombi, with concentration in the outer layers. Moreover, it was found that thrombus NET content correlated significantly with the length of MT procedures and the number of device passes required. This study also expanded upon prior in vitro findings demonstrating that ex vivo, recombinant DNase I accelerated tPA‐induced thrombolysis in NET rich clots.

Additional studies have investigated whether variation in clot NET enrichment is associated with stroke cause (the origin of the clot) or clinical outcomes. In work by Jabrah et al., ³¹ it was found that cardioembolic clots had increased enrichment of neutrophils and extracellular web‐like NETs compared with atherothrombotic or cryptogenic causes. A significantly higher distribution of weblike NETs was also observed around the clot periphery, a finding consistent with prior studies. In another recent work by Lapostolle et al, ³⁰ NET‐rich thrombi were associated with unsuccessful recanalization and longer procedure time, as well as higher rates of postoperative neurological deficit (measured by changes in National Institutes of Health Stroke Scale score) and disability or death (measured by modified Rankin Scale score).

Despite growing recognition of the role of NETs in thrombectomy outcomes, no imaging‐based metric currently exists to estimate clot NET enrichment from preprocedural CT. Developing such a metric is critical for personalized medicine, enabling stratification of patients by their likelihood of benefiting from combination therapy with tPA and DNase I. Noninvasive identification of NET‐rich clots could inform personalized treatment strategies, improve reperfusion rates, and enhance overall thrombectomy success. Therefore, given our transcriptomic results, we chose to further investigate whether clot RFs extracted from prethrombectomy CT could be used to potentiate a noninvasive digital biomarker of NET enrichment and thus first‐pass success. Such radiomics would likely represent either clot NET enrichment or alterations in clot microstructure due to NET enrichment that are reflected on CT.

We first set out to identify radiomics significantly different between mFPE outcomes and found 40 significant RFs. These RFs, computed from CTA and nCCT included several RFs previously reported in the context of AIS. ³⁶ , ³⁷ , ³⁸ Additionally, several of these RFs, such as FOS energy, were identified in prior work as scale‐invariant RFs correlating well with WBC percentage composition. ³³ Given these findings, we next investigated whether this same set of RFs varied with clot NET enrichment. To achieve this, we enlisted existing bioinformatics techniques and databases to derive an estimate of single‐sample gene set enrichment for NET formation. Recognizing that our computed NET enrichment metric was inferred from clot transcriptomes, we further validated our NET score using ground truth immunofluorescence labeling and whole‐slide quantification of CitHis, an established marker of NET formation. ³⁰ , ³¹ , ³² , ³⁹ , ⁴⁰ We suspect that the identified DEGs ontologies indicate a functionally distinct state of neutrophils in clots that did not achieve mFPE, rather than their compositions, because there was no statistically significant difference in percentage of WBCs between the clots of each group (P=0.05) in this data set. Upon further analysis we identified a preliminary 4 RF signature associated with NET enrichment which in cross‐validation achieved a mean AUC of 0.859 in classifying high and low NET clots. Although promising, this RF signature model still needs to be validated in an independent cohort of cases. Overall, the transcriptomic NET metric allowed us to establish a direct link between the thrombus NET enrichment and specific RFs, helping to unravel the intricate interplay between molecular and imaging characteristics in the context of MT first‐pass success.

Our study provides a nuanced understanding of the complex biological factors influencing MT success. Several limitations should be noted. First, the sample size was relatively small, and the study is a single‐center, single‐CT platform design, which may limit the generalizability of our findings. Despite the perfect balance between first‐pass outcomes (n=16 each), the small sample size likely contributes to the identification of only 44 DEGs following correction for false discovery. To explore how the small sample size may have affected the identification of significant features, we performed a post hoc power analysis using the DEGs. Based on the sample size (n=16), the group means, and the population SDs of all the significant genes, and an alpha of 0.05, we found a power 0.87±0.17 for this study. Validation in large, independent, multicenter cohorts with diverse patient populations is needed to confirm the robustness and clinical utility of the identified transcriptomic and radiomic signatures. Second, although our radiomic analysis demonstrates potential for noninvasive prediction of NET enrichment, these findings are based on retrospective imaging data and require prospective validation. In addition, it is important to acknowledge that radiomics associated with NET enrichment may either directly represent clot NET enrichment or represent alterations in clot structure due to increased NET enrichment. Additional NET markers could increase confidence in this study. Lastly, the influence of stroke cause, clot WBC subpopulation composition, clot age, and prehospital treatments on NET‐related signatures warrants further investigation. Nonetheless, our study is among the first to integrate clot transcriptomics with CT radiomics and digital pathology to identify a specific prothrombotic mechanism, NET enrichment, implicated in MT outcomes.

CONCLUSIONS

Our findings highlight the critical role of NET enrichment in influencing first‐pass MT outcomes in acute ischemic stroke. By integrating clot transcriptomics with radiomic analysis of prethrombectomy CT imaging, this is the first study to identify a noninvasive signature of NET‐rich thrombi for stroke. This approach lays the groundwork for stratifying patients before intervention, enabling tailored therapeutic strategies such as the early use of combination tPA‐DNase I therapy or optimized device selection. With further validation in larger, prospective cohorts, radiomic assessment of NET content may become a valuable clinical tool to improve reperfusion rates and overall stroke outcomes.

Sources of Funding

None.

Disclosures

B.A. Santo, T.R. Patel, S.M. Mousavi Janbeh Sarayi, S. Balghonaim, A. Scotti, T.D. Jenkins, V. Jaikumar, and J. Kolega—None. K.E. Poppenberg—Ownership: Neurovascular Diagnostics, Inc. E.I. Levy—Board Membership: Stryker, NeXtGen Biologics, MedX Health, Cognition Medical, EndoStream; Consultancy: Claret Medical, GLG Consulting, Guidepoint, Imperative Care, Medtronic, Rebound Therapeutics, StimMed; Employment: University at Buffalo Neurosurgery Inc; Expert Testimony: renders medical/legal opinions as an expert witness; Stock/Stock Options: NeXtGen Biologics, Cognition Medical, Rapid Medical, Claret Medical, Imperative Care, Rebound Therapeutics, StimMed. A.H. Siddiqui—Financial Interest/Investor/Stock Options/Ownership: Adona Medical, Inc., Amnis Therapeutics, Bend IT Technologies, Ltd., BlinkTBI, Inc, Buffalo Technology Partners, Inc., Cardinal Consultants, LLC, Cerebrotech Medical Systems, Inc, Cerevatech Medical, Inc., Cognition Medical, CVAID Ltd., Endostream Medical, Ltd, Imperative Care, Inc., Instylla, Inc., International Medical Distribution Partners, Launch NY, Inc., NeuroRadial Technologies, Inc., Neurotechnology Investors, Neurovascular Diagnostics, Inc., PerFlow Medical, Ltd., Q’Apel Medical, Inc., QAS.ai, Inc., Radical Catheter Technologies, Inc., Rebound Therapeutics Corp. (Purchased 2019 by Integra Lifesciences, Corp), Rist Neurovascular, Inc. (Purchased 2020 by Medtronic), Sense Diagnostics, Inc., Serenity Medical, Inc., Silk Road Medical, SongBird Therapy, Spinnaker Medical, Inc., StimMed, LLC, Synchron, Inc., Three Rivers Medical, Inc., Truvic Medical, Inc., Tulavi Therapeutics, Inc., Vastrax, LLC, VICIS, Inc., Viseon, Inc. Consultant/Advisory Board: National Center for Advancing Translational Sciences of the National Institutes of Health under award number UL1TR001412 to the University at Buffalo. V.M. Tutino—Financial Interest/Investor/Stock Options/Ownership: Neurovascular Diagnostics, Inc., QAS.ai, Inc. Grant Support: Brain Aneurysm Foundation, National Science Foundation, National Institutes of Health National Institute of Neurological Disorders and Stroke, clinical and translational science institute grant from the national Center for Advancing Translational Sciences of the National Institutes of Health under award number UL1TR001412 to the University at Buffalo. Consultant/Advisory Board: Canon Medical Systems America.

Supporting information

STROBE Checklist

Figures S1–S3