In patients with acute spontaneous intracerebral hemorrhages (ICH), growth of the hematoma volume after hospital admission (“hematoma expansion,” HE) is associated with early clinical deterioration, worse long-term functional outcome, and higher mortality (Brott et al., 1997; Davis et al., 2006; Lord et al., 2015; Hostettler et al., 2019). In the absence of established effective treatments for ICH patients, HE represents a potential therapeutic target (Tanaka and Toyoda, 2021). Identification of patients at elevated risk of HE by means of (imaging) biomarkers or risk scores may allow selective treatment of individuals who likely benefit from anti-expansion therapies in future trials.

In addition to clinical variables (Al-Shahi Salman et al., 2018), the spot sign on admission computed tomography (CT)-angiography (CT-A) has been proposed as a predictor of HE (Demchuk et al., 2012). However, not all centers perform baseline CT-A immediately after identifying an ICH on non-contrast CT, which is the standard-of-care imaging technique for detection of intracranial hemorrhage. Moreover, CT-A is associated with additional ionizing radiation and contrast administration. Alternatively, studies suggested visual markers on non-contrast CT as predictors of ICH expansion (Boulouis et al., 2017; Morotti et al., 2018, 2019). However, overlapping definitions and subjective interpretations of imaging findings limit the applicability and generalizability of such visual markers (Morotti et al., 2019). To date, the clinical value of the CT-A spot sign and visual non-contrast CT markers remains unclear (Hostettler et al., 2019).

A possible alternative is a radiomic biomarker, which allows utilization of standard-of-care non-contrast CTs to provide an objective and reproducible characterization of hematomas (Gillies et al., 2016; Haider et al., 2020a). Radiomic analysis enables a comprehensive, quantitative assessment of shape, density, and texture attributes of volumes-of-interest in medical images through extraction of high-dimensional sets of features (Gillies et al., 2016; Haider et al., 2020a). While the focus of radiomics research thus far were oncological applications (Gillies et al., 2016; Haider et al., 2020a,b,c,d; Tomaszewski and Gillies, 2021), lately stroke radiomics has gained traction (Chen et al., 2021b; Haider et al., 2021; Avery et al., 2022). Recent studies applied radiomic analysis of baseline non-contrast CTs to predict HE; however, with some using small sample sizes, they report a wide range of prediction accuracies (Shen et al., 2018; Xie et al., 2020; Xu et al., 2020; Chen et al., 2021a,c; Pszczolkowski et al., 2021).

Given the need for generalizable imaging biomarkers of HE, which may guide therapeutic interventions in anti-expansion trials, and the equivocal predictive performance of prior radiomic models, we aimed to generate robust non-contrast CT radiomic signatures for HE prediction. Using a large, multicentric dataset of patients prospectively enrolled in the ATACH-2 (Antihypertensive Treatment of Acute Cerebral Hemorrhage II) trial, we devised and independently validated radiomic signatures predictive of ICH expansion. Then, we compared their performance with signatures consisting of visual markers of HE, clinical variables, and combined signatures.

Using a large, multicentric cohort of patients with acute, spontaneous, supratentorial ICH, we devised and validated radiomic signatures for prediction of ICH expansion using features from baseline non-contrast head CT scans. Given that participants were prospectively enrolled in the ATACH-2 trial under controlled conditions, our dataset offers accurate clinical information as well as precisely timed baseline and 24-h follow-up scans enabling rigorous design and validation of HE prediction models. In an independent validation cohort, we demonstrated that a signature combining select radiomic with clinical features of ICH was significantly superior to signatures of visual markers of HE, clinical variables associated with HE, the BAT score and a radiomics-only signature (all p < 0.05). In addition, one should consider the reliability of an automatically extracted radiomic signature versus the complexity of visual assessment of six different HE markers in acute ICH settings. Future studies may combine deep learning hematoma segmentation (Dhar et al., 2020) with radiomics to enable fully automated HE prediction and further reduce reader-dependency. Notably, the fact that neither the BAT score nor visual markers were retained in combined signatures (Supplementary Table 3) suggests that they provide no added predictive value over radiomic features. Moreover, the visual signature yielded numerically but not significantly higher AUCs than the BAT score in both cohorts, suggesting a more comprehensive visual scoring system might yield improved prediction results at the expense of longer and more complex visual image interpretation (Tables 2, 3). Finally, we confirmed the clinical relevance of HE signatures for prognostication of functional outcome by showing consistent associations with 3-month mRS score in the discovery and validation cohorts (Table 4).

Hematoma growth is strongly associated with poor functional outcome and mortality in ICH patients, and therefore, attenuation of ICH expansion is considered a potential treatment strategy (Davis et al., 2006). Unfortunately, thus far, neither intensive blood pressure reduction (Anderson et al., 2013; Qureshi et al., 2016), nor administration of hemostatic drugs such as recombinant factor VII (Mayer et al., 2008) or tranexamic acid (Sprigg et al., 2018; Meretoja et al., 2020), which theoretically target HE, could reduce death or disability in randomized clinical trials. In addition, trials evaluating selective hemostatic therapy of CT-A spot sign-positive patients failed to demonstrate significant treatment benefits (Gladstone et al., 2019; Meretoja et al., 2020). Hence, the search for an effective ICH therapy and (imaging) biomarkers for treatment triage remains ongoing. In this context, an objective and reproducible marker of impending HE based on admission non-contrast head CT−which is readily available and widely used as first-line imaging in emergency departments−may allow future clinical trials to selectively enroll patients who likely benefit from anti-expansion therapy and may ultimately improve ICH outcomes.

In our study, we allocated and strictly separated discovery and independent validation cohorts to accurately quantify radiomic signatures’ performance in predicting HE. Signatures attained very similar AUC scores in both cohorts as well as in cross validation, which is indicative of reliable generalizability. In terms of absolute performance compared to previous studies, our radiomic signature results are similar to those of e.g., Pszczolkowski et al. (2021), who also conducted post hoc analyses of randomized controlled trial data, with an identical HE definition, and similar methodology. On the other hand, Xie et al. (2020), who likewise applied LASSO-LR to devise radiomic signatures, reported AUCs of up to 0.93 in independent validation. The difference in AUC score may be in part attributed to the use of an identical scanner and imaging protocol for all patients by Xie et al. (2020). In addition, the average baseline ICH volume in the study by Xie et al. (2020) was ∼32 ml in patients with HE and ∼12.5 to 14 ml in patients without HE (p < 0.001), suggesting that volume by itself was highly predictive of HE. In our data, however, the baseline volumes differed conspicuously less, with a mean (SD) volume of 13.7 ml (12.7) and 12.2 ml (11.7) among patients with and without HE, respectively (p = 0.13, Wilcoxon rank sum test). As a result, our signatures could not exploit the volume differences. In general, critical appraisal of study populations, methodology, and validation is warranted when comparing radiomics research, where overfitting and information leakage from discovery to validation datasets are frequently encountered challenges.

Multiple visual makers on non-contrast CT were proposed as predictors of ICH expansion (Boulouis et al., 2017; Morotti et al., 2018, 2019). However, overlapping definitions and subjective interpretations may limit their reproducibility. Radiomics, on the other hand, offers reproducible, quantitative, and objective metrics of ICH size, shape, intensity, and heterogeneity characteristics. In this study, we demonstrated that visual markers−alone or in combination−provide no added predictive value to radiomic signatures in prediction of ICH expansion (Supplementary Table 3 and Table 2). In addition, our radiomics-based signatures significantly outperformed the visual signature and BAT score when combined with select clinical predictors. Overall, the objectivity and rapid applicability of radiomic signatures could make them suitable triage tools for multicentric randomized controlled trials, where observer-independent and expeditious enrollment is crucial.

We utilized a large, multicentric, multi-national, prospectively acquired and homogeneous patient dataset with accurately timed baseline and follow-up CT imaging and comprehensive clinical data gathered by a randomized clinical trial under strict oversight, as opposed to previous studies which often relied on retrospective single-center data collection. In addition, we employed state-of-the-art radiomic analysis and strictly separated discovery and validation cohorts to prevent information leakage and performance inflation. The ATACH-2 enrollment criteria, however, inherently limit our findings to patients with acute, spontaneous, supratentorial ICH, hypertension, and a baseline hematoma volume < 60 cm³ (Qureshi et al., 2016). Further studies in more inclusive cohorts are needed to validate our radiomic signatures. Moreover, future studies may compare the predictive value of our signatures with the CT-A spot sign. In addition, although radiomics signatures had significant association with 3-month clinical outcomes, improvement of radiomic HE biomarkers’ absolute predictive performance is crucial before routine clinical application or clinical trials may be considered. It is worth noting that non-contrast head CTs are among the most harmonized medical images: the uniform use of soft tissue kernels, absence of intravenous contrast administration, and calibration of Hounsfield units to exact physical density obviate the need for gray scale normalization. To mitigate the effects of slice thickness and voxel dimension variability, we applied B-spline interpolation to resample images to an isotropic 1 mm × 1 mm × 1 mm voxel spacing, as detailed in the Supplementary methods. To further mitigate the effects of CT slice thickness differences on radiomic feature values, we applied ComBat harmonization. However, while achieving a slight numeric improvement in prediction accuracy within the discovery cohort, signatures compiled from harmonized radiomic features yielded numerically inferior AUCs in the validation cohort, which may be indicative of overfitting (compare Table 2 and Supplementary Table 5). There is still ample potential for refinements, which may include further harmonizing CT imaging protocols across centers, usage of higher resolution scans and reconstructions, automated segmentation algorithms and incorporation of radiomic features from additional ICH manifestations such as the perilesional edema or intraventricular hemorrhage. Nevertheless, we believe our study, confirming the results of some prior reports, underlines the value of radiomics in HE prediction.

Using a large multicentric dataset, we generated and independently validated a radiomic signature of HE based on admission non-contrast head CTs of patients with supratentorial ICH. We demonstrated that a signature combining radiomic features and clinical predictors significantly outperforms a signature of visual CT markers of HE as well as the BAT score, and that adding visual markers to radiomic features offers no improvement in predictive performance. All HE signatures were significantly associated with 3-month functional outcome, underlining their prognostic relevance. Limited to ICH patients with similar characteristics, the proposed markers may enable selective anti-expansion treatment of patients at higher risk of HE in future clinical trials.

The data analyzed in this study is subject to the following licenses/restrictions: data are available on reasonable request, and on approval from the respective register holders. Requests to access these datasets should be directed to the ATACH-2 investigators (ClinicalTrials.gov identifier: NCT01176565).

Ethical compliance was ensured by the ATACH-2 investigators (ClinicalTrials.gov identifier: NCT01176565). Our group performed post hoc analyses of anonymized data. The patients/participants provided their written informed consent to participate in this study.

SH: conceptualization, data curation, formal analysis, investigation, methodology, project administration, software, validation, visualization, writing—original draft, writing—review and editing. AQ, AJ, HT, and EB: data curation, investigation writing—review and editing. TZ: formal analysis, investigation, writing—review and editing. DW, MG, AdM, AjM, LS, GF, and KS: investigation, writing—review and editing. SP: conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, resources, software, supervision, validation, writing—original draft, writing—review and editing. All authors contributed to the article and approved the submitted version.