^{1}

^{1}

^{1}

^{1}

^{2}

^{3}

^{1}

^{*}

^{1}

^{2}

^{3}

Edited by: Baojuan Li, Fourth Military Medical University, China

Reviewed by: Xia-an Bi, Hunan Normal University, China; Xiaopan Xu, Fourth Military Medical University, China

This article was submitted to Brain Imaging Methods, a section of the journal Frontiers in Neuroscience

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

We aimed to construct and validate a nomogram model based on the combination of radiomic features and satellite sign number for predicting intracerebral hematoma expansion.

A total of 129 patients from two institutions were enrolled in this study. The preprocessed initial CT images were used for radiomic feature extraction. The ANOVA-Kruskal–Wallis test and least absolute shrinkage and selection operator regression were applied to identify candidate radiomic features and construct the Radscore. A nomogram model was developed by integrating the Radscore with a satellite sign number. The discrimination performance of the proposed model was evaluated by receiver operating characteristic (ROC) analysis, and the predictive accuracy was assessed

Four optimal features were ultimately selected and contributed to the Radscore construction. A positive correlation was observed between the satellite sign number and Radscore (Pearson’s

A nomogram model of integrated radiomic signature and satellite sign number based on noncontrast CT images could serve as a reliable and convenient measurement of hematoma expansion prediction.

Intracerebral hemorrhage (ICH) confers a worse prognosis than ischemic stroke, with an overall fatality rate approaching 40% and neurological disability among the survivors (

Recently, different imaging characteristics have been successively reported and paved the way for available prediction of hematoma expansion in clinical routine. The computed tomography angiography (CTA) spot sign, as an independent predictor, has been well established and prospectively validated, which turned out to be of limited sensitivity (

In this study, we hypothesized that radiomic analysis and quantitative satellite sign can identify the associations between the quantitative imaging features and the hematoma pathophysiology and thus effectively and precisely predict intracerebral hematoma expansion in NCCT images. The aim of this study was to establish a quantitative imaging model to predict hematoma expansion and improve the functional outcomes for patients with ICH. We investigated a nomogram model combined with radiomics and quantitative satellite sign to improve the diagnostic performance in early hematoma expansion prediction.

This retrospective study was approved by the Medical Ethics Committee of institution I and II and conducted in accordance with relevant guidelines. Informed consent was waived.

Patients with spontaneous ICH within 6 h since symptom onset and CT recheck within 24 h in between January 2017 and December 2018 were included. The exclusion criteria were the following conditions: (1) patients with ICH secondary to arteriovenous malformation, trauma, aneurysm, tumor, and venous sinus embolism, (2) patients who were receiving anticoagulation treatment, (3) surgery or interventional therapy before the repeat CT scan, (4) image contained severe artifacts, and (5) IVH or subarachnoid hemorrhage is involved. Clinical data were provided by a neurologist, including age, gender, systolic blood pressure, international normalized ratio, time to initial CT scan, activated partial thromboplastin time, and baseline Glasgow Coma Scale score.

The CT scans in the two institutions were carried out on different CT scanners, including a GE LightSpeed VCT 64-slice and a GE Optima 540 16-slice. The same CT scanning parameters were performed with a tube voltage of 120 kV, a tube current of 150–300 mA, field of view of 25 cm, and 512 × 512 acquired matrix. The scan ranged from the skull base to the cranium, with a thickness of 5 mm per layer.

The radiomic workflow is summarized in

Workflow.

All the radiomic features from ROIs were extracted from preprocessed images using the Artificial Intelligence Kit Version 3.0.1.A (Life Sciences, GE Healthcare, United States), with window width 110 and window level 45. Six main categories were involved, including histogram, morphology, gray level co-occurrence matrix, run length matrix, and gray level zone size matrix. Analysis of variance, Kruskal–Wallis test, and single-factor logistic regression analysis were successively carried out for selecting significant features that were highly correlated. By removing the redundancy with a correlation coefficient of more than 0.90, the radiomic features were further optimally elected. In the final step, least absolute shrinkage and selection operator (LASSO) regression was applied to identify the most nonredundant and robust features among the 396 radiomic features from the training cohort in order to improve the class separability and optimize the representation of lesion heterogeneity. With an increase of the value of

Both Radscore and the satellite sign number were integrated by a multivariate logistic regression-based radiomic model in the training cohort. Furthermore, a nomogram model was constructed based on a multivariate logistic regression analysis to visually demonstrate the probability of a hematoma enlargement. In addition, predictive models based on Radscore or the satellite sign number alone were also developed. The receiver operating characteristic (ROC) analysis and the AUC were applied to evaluate the discrimination performance on the three models. Along with the Hosmer–Lemeshow test measuring for goodness of fit of the nomogram model, predictive accuracy was assessed

Decision curve analysis (DCA) was carried out to evaluate the clinical value of the three models independently on the basis of calculating the net benefit for patients at each threshold probability. By comparing to all strategies or none at all, the best model was elected according to the higher calculated net benefit.

The Kaplan–Meier method was carried out to calculate the survival probabilities. The survival rates were estimated in 30 days. The patients from the two institutions were divided into the expander and the non-expander groups according to the predictive results using the threshold calculated from the training dataset through the Youden Index. Survival was defined as the period from diagnosis to the date of death or the time at which information was last obtained.

Version 3.3.2 of R software and version 13.0 of SPSS software were used in the statistical analysis. Quantitative variables are shown as mean ± SD. Statistical group comparisons of clinical data were performed by independent-samples ^{2} test where appropriate. Intraclass correlation coefficient (ICC) was analyzed for estimating the reliability of inter-observer agreements, which was defined as good consistency if between 0.75 and 1, fair consistency if between 0.4 and 0.75, and poor consistency if under 0.4. The pairwise comparison of ROC curves was performed using z statistic in MedCalc for Windows, version 19.0.7 (MedCalc Software, Ostend, Belgium). Log-rank test was used to compare survival curves, and the results were considered as significant when ^{∗}Power software. The level of statistical significance was set at a two-sided

As demonstrated in the workflow (

Baseline demographic information.

Variable | Training set ( |
Validation set ( |
||||

Expander ( |
Non-expander ( |
Expander ( |
Non-expander ( |
|||

Age (years) | 64.2 ± 14.7 | 61.2 ± 14.2 | 0.42 | 56.2 ± 10.6 | 57.1 ± 13.5 | 0.86 |

Male (%)* | 13 (19.1) | 31 (45.6) | 0.75 | 6 (20.7) | 12 (41.4) | 0.98 |

Admission SBP (mmHg) | 168.5 ± 8.8 | 165.0 ± 11.2 | 0.21 | 162.3 ± 9.1 | 166.1 ± 8.2 | 0.28 |

Admission INR | 1.5 ± 0.2 | 1.5 ± 0.3 | 0.25 | 1.4 ± 0.3 | 1.5 ± 0.3 | 0.51 |

Time to initial CT scan (h) | 3.1 ± 0.9 | 3.5 ± 1.1 | 0.21 | 3.9 ± 1.6 | 3.9 ± 1.8 | 0.96 |

APTT (s) | 33.3 ± 4.9 | 32.0 ± 3.9 | 0.25 | 28.3 ± 6.5 | 29.5 ± 5.0 | 0.62 |

Baseline GCS score | 12.2 ± 3.8 | 12.0 ± 3.4 | 0.84 | 13.7 ± 3.7 | 12.3 ± 3.8 | 0.37 |

^{∗}Data are the number of patients, with percentages in parentheses.

Radiological characteristics.

Variable | Training cohort ( |
Validation cohort ( |
||||

Expander ( |
Non-expander ( |
Expander ( |
Non-expander ( |
|||

Basal ganglia | 16 (76.2) | 35 (74.5) | 0.88 | 11 (57.9) | 31 (73.8) | 0.21 |

Lobar | 4 (19.1) | 9 (19.2) | 0.99 | 5 (26.3) | 7 (16.7) | 0.38 |

Thalamus or brainstem | 1 (4.8) | 3 (6.4) | 0.79 | 3 (15.8) | 4 (9.5) | 0.67 |

Satellite sign number | 2.4 ± 1.6 | 1.0 ± 1.5 | <0.001 | 2.4 ± 1.7 | 0.7 ± 1.0 | 0.001 |

Black hole sign* | 8 (38.1) | 9 (19.1) | 0.10 | 7 (36.8) | 7 (16.7) | 0.08 |

Swirl sign* | 7 (33.3) | 6 (12.8) | 0.04 | 8 (42.1) | 9 (21.4) | 0.09 |

Blend sign* | 9 (42.9) | 10 (21.3) | 0.07 | 7 (36.8) | 9 (21.4) | 0.21 |

Basal ganglia | 16 (76.2) | 36 (76.6) | 0.97 | 11 (57.9) | 31 (73.8) | 0.21 |

Lobar | 4 (19.1) | 8 (17.0) | 0.84 | 5 (26.3) | 6 (14.3) | 0.23 |

Thalamus or brainstem | 1 (4.8) | 3 (6.4) | 0.79 | 3 (15.8) | 5 (11.9) | 0.69 |

Satellite sign number | 2.5 ± 1.7 | 1.1 ± 1.4 | 0.001 | 2.2 ± 1.1 | 0.7 ± 0.6 | 0.003 |

Black hole sign* | 8 (38.1) | 8 (17.0) | 0.06 | 9 (47.4) | 10 (23.8) | 0.07 |

Swirl sign* | 7 (33.3) | 7 (14.9) | 0.08 | 8 (42.1) | 10 (23.8) | 0.15 |

Blend sign* | 7 (33.3) | 8 (17.0) | 0.13 | 7 (36.8) | 8 (19.0) | 0.14 |

^{∗}Data are the number of patients, with percentages in parentheses.

Based on the result of the reproducibility analysis by the two radiologists, 349 out of 396 (88.2%) radiomic features had good consistency (ICC ≥ 0.75) on contour-focused segmentation. The number of features with fair consistency (0.75 > ICC ≥ 0.4) and with poor consistency (ICC < 0.4) was 28 (7.1%) and 19 (4.7%), respectively.

Univariate analysis of four candidate features for hematoma expansion prediction in the training cohort. HGLRE, high gray level run emphasis, SRHGLE, short run high gray level emphasis. *

From the pairwise Pearson correlative analysis, the satellite sign number was observed to be positively correlated to the corresponding Radscore with a correlation coefficient of 0.482 (

Correlation between satellite sign number and Radscore based on Pearson correlation analysis. The mean absolute correlation was 0.482 (

Nomogram construction and performance of the combined model in both cohorts.

A further validation was carried out through ROC analysis. Compared to the Radscore (0.812, 95% CI: 0.698 to 0.897, sensitivity: 0.992, specificity: 0.617) and the satellite sign number (0.762, 95% CI: 0.643 to 0.858, sensitivity: 0.950, specificity: 0.511) alone, the combination of the two yielded an even better performance in the prediction of hematoma expansion as well as an increased AUC of 0.881 (95% CI: 0.779–0.947, sensitivity: 0.973, specificity: 0.787) in the training cohort (

According to the ROC analysis, the nomogram model yielded a higher AUC value (0.857, 95% CI: 0.750–0.931, sensitivity: 0.950, specificity: 0.766) than the Radscore-based model (0.776, 95% CI: 0.657 to 0.868, sensitivity: 0.750, specificity: 0.745) and the satellite sign number (0.720, 95% CI: 0.597 to 0.823, sensitivity: 0.950, specificity: 0.426) in the external validation cohort. Consistent results were shown in the pairwise comparison of ROC curves (

DCA was conducted to assess the clinical utility of the nomogram model (

Decision curve analysis for the nomogram model in the external validation cohort. The gray line stands for the assumption that all patients developed hematoma expansion, and the black line represents the assumption that no patient had hematoma expansion. Compared to other models, the highest curve of the nomogram model with more area is the optimal decision making for maximal net benefit in hematoma expansion prediction.

The Kaplan–Meier survival analysis showed approximate survival rates between actual subjects and predicted ones. Furthermore, a significant difference was found not only between the actual expander and non-expander groups but also between predicted groups, which suggested the prognostic value of the combined nomogram model (

Kaplan–Meier (KM) survival curve for actual and predicted expander and non-expander groups. The KM analysis shows a significant difference between both actual and predicted groups (

In this study, we established and validated a nomogram model for early ICH expansion prediction, incorporating four robust radiomic features which were extracted from NCCT and proven to be effective for the classification of expanders and non-expanders and the satellite sign number which was found to be a statistically significant imaging marker for the identification of expanders. The nomogram model achieved a significantly better performance in both training cohort and external validation cohort with a larger AUC value than the model of radiomic signature alone, suggesting the reproducibility and the reliability of the improved model in hematoma expansion prediction.

In the recent years, several imaging markers for assessing the greater risk of ICH expansion in NCCT images have been springing up (

The satellite sign and the island sign shared a similar morphology-based definition, which was less influenced by photon noise during observation (

The LASSO regression method has already been widely applied in radiomics-based studies (

Early hematoma expansion is a critical determinant for both mortality and dependency after ICH onset. As the only modifiable factor in the vast majority of patients, it takes the center stage in therapeutic strategies (

There were some limitations in the current research that still need to be further investigated. First of all, it was a retrospective study with a relatively small and imbalanced sample size between expanders and non-expanders in both training and external validation cohorts. Further prospective researches are warranted to expand and balance the sample size and to verify the conclusions. On the other hand, in the process of hematoma segmentation, when it comes to hematoma located in cortical or subcortical regions, it was prone to inaccurate delineation due to partial volume effects. Besides that, the feature extraction software made the displacement vectors 1, 4, and 7 describe the relationship between the gray scale of pixels of the texture as default setting. In light of this, different set points could possibly influence the quantity and the category of radiomic feature extraction; thus, a future radiomic analysis based on various displacement vectors is required. Due to the relatively short follow-up time, the median overall survival for ICH was not available. We will continue to follow up with these patients to secure a more complete prognosis status.

We have identified and validated a nomogram model of integrated radiomic signature with the satellite sign number based on NCCT images to be a reliable and precise evaluation measurement for ICH enlargement prediction at early baseline. The predictive model could serve as an objective and convenient tool to use for patients with ICH in individualized prediction and treatment decision-making, thus suggesting a great potential for clinical application.

All datasets generated for this study are included in the article/

The studies involving human participants were reviewed and approved by Affiliated Hangzhou First People’s Hospital, Zhejiang University School of Medicine. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

WX and QS wrote the manuscript and YS and ZD contributed to the writing process. WX, PP, ZF, and QS analyzed and interpreted the data and prepared the tables and figures. WC, YS, and ZD acquired the data. QS additionally contributed to the conception and the design of the study. All the co-authors read and revised the article.

PP was employed by company GE Healthcare (China). The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The Supplementary Material for this article can be found online at: