Research advances in predicting the expansion of hypertensive intracerebral hemorrhage based on CT images: an overview

Introduction

Intracerebral hemorrhage (ICH) is the second most common type of stroke, accounting for 10–20% of all strokes, and is associated with high morbidity and mortality (Schrag & Kirshner, 2020; Sheth, 2022; Vandertop, Can & Post, 2023). Hypertensive intracerebral hemorrhage (HICH) is the most common type of ICH, accounting for about 50–70%. With the aggravation of the population aging, the incidence of HICH is increasing year by year (Anand et al., 2023; Hostettler, Seiffge & Werring, 2019). Hematoma expansion (HE) is considered to be the mortality or poor prognostic factors after ICH, and its incidence is about 13–38% (Chu et al., 2023; Li et al., 2020). According to reports in the literature, for each 1 ml increase in hematoma after ICH, which was associated with a 5% increase in the hazard of death or dependency (95% confidence interval [CI], [2∼9%]). Conversely, any intervention that reduces hematoma growth by 2 to 4 ml is equivalent to a 10% to 20% reduction in the risk of death or dependency within 90 days (Chu et al., 2022; Delcourt et al., 2012). In clinical practice, accurately predicting the risk of HE is helpful for early targeted treatment or surgery (Al-Shahi Salman et al., 2018), which can prevent adverse events, reduce patient mortality and disability rates, and have important clinical significance (Limin et al., 2022).

How to predict HE early, accurately, and noninvasively has become a hotspot and challenging issue in current clinical research (Hu et al., 2021; Puy et al., 2023). Computer tomography (CT) is the preferred examination method for the diagnosis and follow-up of acute HICH (Nehme et al., 2022), including CT plain scan and computer tomography angiography (CTA) (Rindler et al., 2020). It has important value in the diagnosis, localization, and hematoma quantification of HICH (Romero & Rojas-Serrano, 2023). In the past, many studies and clinical practice had proved that some CT imaging signs are independent predictors of HE, such as “blend sign”, “black hole sign”, “island sign” and “swirl sign” (Al-Ajlan et al., 2023; Huang et al., 2023b). However, the above signs are not common in clinical practice. Moreover, relying solely on CT signs to predict hematoma enlargement has the drawbacks of strong subjectivity, high dependence on the diagnostic level of doctors, and poor consistency. Therefore, further research is needed on how to more effectively utilize CT images to accurately assess HE risk.

In recent years, radiomics and artificial intelligence have also shown strong application prospects in the integration of imaging, genetic, clinical, and other information to assist diagnosis, analysis, and prediction (Chen et al., 2023, 2022). The combination of artificial intelligence and medicine has had a profound impact on the medical system and greatly expanded the value of medical imaging in clinical practice (Wang et al., 2021). At present, it has been widely used in the diagnosis and prognosis evaluation of mild cognitive impairment, Alzheimer’s disease, Parkinson’s disease, lung cancer, liver cancer, breast cancer and other diseases (Feng et al., 2021; Feng & Li, 2023; Georgiadou et al., 2023). With the rise of radiomics and artificial intelligence, significant progress has been made in predicting HE (Zhao et al., 2022).

The rationale of our work is based on a detailed investigation of methods for predicting intracranial hematoma enlargement using CT images. We have discovered that many previous studies on this topic may not have captured the information we obtained from various authoritative academic data sources. Previous efforts to expand recognition of intracranial hematoma based on CT images have been limited to a small number of academic sources. Some early studies failed to adequately emphasize the advantages and disadvantages of different methods for predicting intracranial hematoma enlargement using CT images. The innovation in our work lies in identifying targets for predicting intracranial hematoma enlargement based on CT images, utilizing data from multiple reputable academic sources. Additionally, we highlight important strategies along with their respective advantages and disadvantages in predicting intracranial hematoma enlargement. Therefore, this article reviews the latest research progress in predicting HE based on CT imaging, radiomics, and artificial intelligence, in order to assist in early clinical identification of high-risk HE patients.

Survey methodology

This study strictly adhered to the norms of a systematic literature review and carefully designed research methods to ensure the integrity and rigor of the investigation. Specifically, we proceeded with our planned data collection and conducted a thorough review of the data.

In the first phase of data collection, we conducted a comprehensive search of Pubmed, Springer, Science Direct, and Web of Science to find academic articles relevant to our study’s prediction of intracranial hematoma enlargement. The initial search aimed to cover a wide range of literature on the subject. After an extensive search, we adopted a more targeted approach. Specifically, we have conducted an exhaustive study of literature that is widely cited and directly related to our research topic. We have applied rigorous methodologies and outstanding features to ensure a thorough, comprehensive review. By combining practical insights from academic research, our aim is to explore and present the complex issues involved in predicting the expansion of hematoma in the brain from multiple perspectives. It is hoped that this work can provide a scientific basis for formulating treatment plans early on and make a positive contribution to improving patients’ quality of life.

This study was approved by the ethics committees of Chongqing University Central Hospital (Chongqing Emergency Medical Center). Due to the retrospective nature of this study, the ethics committees waived the need for individual consent.

The ct plain scan imaging signs in prediction of he

Irregular hematoma shape and heterogeneous hematoma density

In 2009, Barras et al. (2009) reported that irregular shape and heterogeneous density of hematoma were independent risk factors for HE, and formulated a five-point categorical scale (Fig. 1). The hematoma shape score: 1 score represented that the hematoma had a regular shape, no lobes or satellite lesions, 2 scores indicated that the hematoma had a lobe or satellite lesion, and so on. The highest score was five points, and the higher the score, the greater the possibility of HE occurrence. The 1–2 points were defined as regular shapes of hematoma, and the 3–5 points were defined as irregular shapes of hematoma. The hematoma density score: 1 score indicated uniform density or one low-density area of the hematoma, 2 scores represented two low-density areas in the hematoma, and so on. The highest score was five points, and the higher the score, the greater the risk of HE. The 1–2 points were defined as homogeneous hematoma density, and the 3–5 points were defined as heterogeneous hematoma density. Yu et al. (2017) reported in the meta-analysis that the sensitivity and specificity of irregular hematoma shapes in predicting HE were 67% and 47%, respectively. And the sensitivity and specificity of hematoma density heterogeneity in predicting HE were 52% and 69%, respectively. The irregular shape and heterogeneous density of the hematoma indirectly reflect the existence of perihematoma injury and active bleeding in the hematoma (Liu et al., 2022). However, the sensitivity and specificity are low in the prediction of HE, which is not conducive to clinical promotion.

Hypodensities

In 2016, Boulouis et al. (2016b) proposed that the hypodensity sign predicted HE, which was defined as the existence of relatively low-density areas in the hematoma that were not connected with the surrounding brain parenchyma (Fig. 2). Vedartham et al. (2021) found that the detection rate of the hypodensity sign was related to the time of onset. The CT detection rate of patients within 3 h after symptom onset (about 71%) was higher than that after 3 h (about 29%), and the detection rate declined with the extension time. Boulouis et al. (2016a) reported that the incidence of hypodensities with the hematoma was 31.2% overall, 52.8% in the HE group, and 21.9% in the non-HE group (HE group vs. non-HE group P < 0.001), and it was significantly correlated with HE (odds ratio, OR, 3.42 95%CI [2.21–5.31]), which could be used as an independent predictor of HE.

Blend sign

In 2015, Li et al. (2015) found that the blend sign could effectively predict HE (Fig. 3). The blend sign was characterized by hypoattenuating and hyperattenuating regions within the hematoma, with a well-defined margin, and the difference in CT value between the two density regions was ≥18HU. The incidence of the blend sign was 16.9%, with a sensitivity of 39.3% and a specificity of 95.5% for predicting HE. Zhang et al. (2020) compared the effectiveness of the blend sign and other CT plain scan signs in predicting HE and found that the blend sign was better than the black hole sign in predicting HE (AUC: 0.67 vs. 0.57; P < 0.001). This imaging marker could help identify patients at a high risk of hematoma expansion and aid in their early treatment (Li et al., 2021; Schreiber et al., 2023).

Black hole sign

Subsequently, Li et al. (2016) reported that the black hole sign predicted HE (Fig. 4). The black hole sign represents a well-defined hypodense region within a hyperdense hematoma that is not connected to the nearby brain parenchyma. It could be round, oval, or rodlike. Moreover, it was clearly separated from the hyperattenuating region of the hematoma, and the CT value difference between the two was ≥28HU.The incidence of the black hole sign was 14.6%. The sensitivity and specificity of predicting HE were 31.9% and 73.2%, respectively. Chen et al. (2020) conducted a meta-analysis that showed the black hole sign predicts HE with a 30% sensitivity and 93% specificity, and the area under the cure (AUC) was 0.83. As an independent predictor of HE, although the black hole sign has a high specificity, its sensitivity is low, and its clinical application is limited (Shakya et al., 2021).

Swirl sign

The swirl sign denotes one or more hypodensity areas in the hyperdensity hematoma (Fig. 5), which morphology is variable, round, streaklike, and strip. And the hypodensity area can be continuous with the adjacent brain parenchyma (Ng et al., 2018; Park et al., 2019). Amoo et al. (2021) found in a meta-analysis study evaluating nine articles that the swirl predicts HE with 50% sensitivity and 77% specificity. Yu et al. (2019a) conducted a meta-analysis to assess the accuracy of swirling signs in predicting intracerebral hematoma enlargement. A total of 2,647 patients from six studies were included in the analysis. The results revealed that the sensitivity and specificity of swirling signs were 0.45 (95%CI [0.32–0.59]) and 0.79 (95%CI [0.73–0.84]), respectively, while the combined positive likelihood ratio was found to be 2.2 (95%CI [1.8–2.5]). Additionally, the combined negative likelihood ratio for the vortex sign was determined as 0.69 (95%CI [0.57∼0.84]). This study suggests that swirling signs exhibit high specificity in predicting hematoma enlargement among patients with cerebral hemorrhage.

Island sign

In 2017, Li et al. (2017) discovered and proposed the island sign again (Fig. 6), which has been confirmed to be an important imaging marker of early HE and poor clinical prognosis. The island sign is defined as ≥3 small discrete hematomas that separated from the main hematom or ≥4 small hematomas, which may branch off from or be completely connected with the main hematoma, should be round, vesicular, or budlike but not lobulated. Zhou et al. (2021) conducted a meta-analysis study on island sign, which showed that the comprehensive sensitivity and specificity of island sign in predicting HE were 50% and 89%, respectively, and its AUC was 0.73. However, the occurrence of the island sign is low, about 21%, and its clinical use is limited (Wei et al., 2020).

Satellite sign

The satellite sign is defined as a small hemorrhagic focus that surrounds the main hematoma (Fig. 7), no more than 1 cm in maximum diameter, 1 to 2 cm from the main hematoma, and not within the ventricles and subarachnoid space (Lv et al., 2021; Songsaeng et al., 2023). In a meta-analysis, the satellite sign has been shown to predict HE with a sensitivity of 50% and a specificity of 71% (Yang et al., 2020), which was an effective marker for predicting HE. The incidence of the satellite sign was 49.2%, which was significantly higher than that of the island sign. The incidence of HE with the positive satellite sign was 40.4%, and the incidence of HE with the positive island sign was 91.3%.

Fluid level sign

The fluid level is a stratified phenomenon in the cerebral parenchyma (Fig. 8), with areas of low CT attenuation above and high CT attenuation below a discrete line of separation in an area of intraparenchymal hemorrhage, which is rarely seen in ICH patients and lacks validation (Morotti et al., 2023). At present, there are few studies on this sign, and it is included in the scoring scale as a predictor in some HE prediction scales (Mendes Pedro et al., 2021).

The overlapping CT plain scan imaging signs

There are some overlap in the attenuation between the imaging signs of hematoma density, such as the swirl sign, fluid level sign, black hole sign, and the hypodensities in the hematoma that is not connected with the adjacent brain parenchyma (Fig. 9) (Chung et al., 2022; Nawabi et al., 2023). In the fluid level sign, a hematoma with an attenuating difference of ≥18HU between two areas of differing densities is also the blend sign. The density of a hematoma is considered to be homogeneous when there are 1–2 imaging signs of density characteristics (Barras score I–II) (Delcourt et al., 2016). When more than three imaging signs of density features are present, the density of hematoma is considered heterogeneous (Barras score III–V). There are also some overlap in the attenuation between the imaging signs of the morphology of hematoma. When the number of the satellite sign is ≥3, it is also the island sign (Deng et al., 2019). Regular morphology (Barras score II) is defined when there is only one satellite. The presence of two or more satellite signs or island signs is defined as irregular hematoma shape (Barras score III–V) (Yu et al., 2019b).

The imaging signs asscociated with he on cta

CTA spot sign

The CTA spot sign was first proposed by Wada et al. (2007) in 2007 (Fig. 10). The spot sign is defined as one or more foci with a diameter of 1–2 cm of enhancement within the hematoma on CTA source images. It has been confirmed in a number of studies and clinical work that it is closely related to the occurrence of HE, and the spot sign can predict HE with 91% sensitivity, 89% specificity, 77% positive predictive value and 96% negative predictive value. Subsequently, many research studies have modified the definition of the CTA point sign, and now the following definition is mostly used (Dowlatshahi et al., 2016; Romero et al., 2012): ① There is ≥1 contrast leakage from the hematoma. ② The CT value of spot foci is at least 120HU higher than that of peripheral hematoma. ③ The spot foci is discontinuous with any adjacent vessels. ④ The spot foci could be of any size and shape, and the above four criteria should be met at the same time. In the research, the inconsistent use of spot sign definitional criteria led to inconsistent conclusions on its predictive value for HE. In 2018, Xu et al. (2018) included a total of 29 studies in a meta-analysis on CTA prediction of HE, and found that the incidence of spot signs was 23.4%, and the predictive sensitivity and specificity of HE were 62% (95% CI, [54–69%]) and 88% (95% CI, [85–91%]), and the OR of spot signs with HE was 8.49 (95% CI, [7.28–9.90]). At present, CTA spot sign is recognized as one of the most effective imaging markers for predicting HE, which shows the contrast agent extravasation from the microvessels and represents the presence of active bleeding. Meanwhile, the presence of spot sign is also a risk factor for in-hospital death and 3-month death of ICH patients, as well as an essential predictor of poor prognosis at discharge and 3 months of patients (Singh et al., 2021).

Predicting he with ct radiomics

Overview of radiomics

Radiomics is a feature set which high-throughput converts medical images into the data features, and then quantitative analysia of these data features, which offer the diagnose and predict efficacy and prognosis (Mayerhoefer et al., 2020). Radiomics is a natural extension of computer-aided diagnosis and detection system, which has the advantages of non-invasiveness, repeatability, multi-level and even the microscopic characteristics of the whole lesion data mining and comprehensive evaluation, the ability to interpret the microscopic characteristics of the lesion from the overall spatial heterogeneity, and can play the role of virtual biopsy, ect (Huang et al., 2023a). Radiomics usually involves the following four workflows (Fig. 11): (1) Imaging image acquisition, including CT, MRI, and ultrasound images. (2) ROI drawing and image segmentation. (3) Feature extraction and selection: the data features for quantitative analysis of lesion heterogeneity, including morphological characteristics, first-order statistical characteristics, second-order statistical characteristics, and higher-order statistical characteristics, were extracted from the segmented images. Due to the large number of radiomics features extracted, even up to thousands of different features, too many dimensions will cause the risk of model overfitting, so it is particularly important to carry out feature reduction and selection (Huang et al., 2023a). Common methods include least absolute shrinkage and selection operator regression algorithm, principal-component analysis, etc. (4) Model construction and validation: the extracted and screened features are used to construct prediction models, and Logistic regression analysis is the most common method. In recent years, machine learning algorithms have also been widely used in model construction and have obtained greater discrimination ability. Finally, assessing of the radiomics modle, which include internal and/or external verification, as well as ROC curve, calibration curve and clinical decision curve.

Research progress of artificial intelligence based on deep learning in he prediction

Discussion

HICH is a complex and dynamic process (Chiquete et al., 2007). A large amount of blood can compress and damage surrounding tissues, leading to impaired nerve function. Patients may experience symptoms such as headaches, changes in consciousness, and movement disorders (Weimar & Kleine-Borgmann, 2017). If the hematoma expands, it often indicates a deterioration of neurological function and an increase in mortality rate. Therefore, early and accurate identification or prediction of HE is beneficial for developing personalized treatment plans and improving patient prognosis (Romero & Rojas-Serrano, 2023).

With the development of medical technology, various methods have shown important potential in predicting HE. CT plain scan has the advantages of simple operation and fast examination speed, which has been widely used in clinical practice (Ma & Zhou, 2022). It can provide a large amount of lesion information, including bleeding volume, density, location, and displacement of surrounding structures. By accurately obtaining and analyzing this information, it can help predict whether the hematoma will further expand. The irregular shape and heterogeneous density of hematoma are considered as risk factors for HE (Liu et al., 2022). However, they have relatively low specificity, which limits their clinical application. The specificitiy of characteristics including mixed sign, black hole sign, and island sign are higher, but their incidence is low (Gupta & Seth, 2019). Meanwhile, there may be overlapping features on CT images, which poses a certain challenge for radiologists. The dot sign on CTA is currently recognized as one of the most effective imaging markers for HE (Kim et al., 2019). Its appearance suggests extravasation of contrast agent and active bleeding, which has high sensitivity and specificity. However, due to the need for intravenous contrast agents in CTA, there is risks of allergies and renal dysfunction for this examination. Therefore, it is difficult to predict HE by conventional CT imaging features, even for experienced radiologists.

Radiomics is a newly emerging computer-aided technology that can quantify the internal information of hematoma by high-throughput extraction of massive features from medical images (Dai et al., 2023). This technology is low-cost, non-invasive, simple, and repeatable, without the need for additional scanning equipment, which has broad application prospects. At present, various radiomics models have been established to predict hematoma enlargement, demonstrating good accuracy and clinical value (Li et al., 2022). Deep learning, such as convolutional neural networks and stacked autoencoder networks, can automatically learn a large amount of data features and achieve high-precision prediction and classification through continuous optimization. Deep learning algorithms can continuously adjust parameters to adapt to different scenarios and data structures. According to literature reports, the accuracy of deep learning models in predicting hematoma enlargement exceeds that of traditional clinical variable models and machine learning models (Zhao et al., 2023). However, radiomics and deep learning are still in their early stages in this field, lacking real-world research, and some models may not perform well when applied to external data. In future research, real-world randomized controlled trials should be conducted. On the other hand, the knowledge of radiology experts with clinical experience should be introduced into artificial intelligence to achieve a closed loop between artificial intelligence and human intelligence.

Conclusions

HE is an important risk factor for neurological deterioration and death in HICH patients. Accurately predicting the risk of HE is of great significance for achieving precise intervention and improving patient prognosis. Previous studies have found that many imaging features have important clinical value in predicting HE. However, cerebral hemorrhage is a complex and dynamic process, and traditional CT imaging features have a certain degree of subjectivity and poor predictive accuracy. The rise of radiomics and artificial intelligence technology has provided a new method for quantitative research on HE, significantly improving the accuracy of HE prediction. However, current artificial intelligence methods face the problem of long time for delineating regions of interest and model building. In the future, artificial intelligence technology should solve this problem and achieve fast and automated prediction. Once this issue is resolved, it will achieve widespread clinical application and promotion, which has important clinical significance.