Stroke is a devastating condition associated with high risks of disability, mortality, and recurrence, posing a significant socioeconomic burden (Liu et al., 2020). Over the past several decades, the global burden of stroke has continued to rise, driven by population growth, aging demographics, and the increasing prevalence of risk factors (Katan and Luft, 2018).

Hemorrhagic transformation (HT) after ischemic stroke is defined as the detection of bleeding on a follow-up computed tomography (CT) or magnetic resonance imaging (MRI) scan in patients with no evidence of hemorrhage on the initial post-infarction scan, or hemorrhagic infarction identifable on the initial scan (Chinese Society of Neurology Chinese Stroke Society, 2019). HT is a critical concern in the management of acute ischemic stroke and represents part of the natural history of cerebral infarction (Moulin et al., 1994). The clinical incidence of HT ranges from 2.2% to 44.0%, while the pathological incidence can reach up to 70% (Aviv et al., 2009). Current treatment options for ischemic stroke, including intravenous thrombolysis, endovascular therapy, anticoagulants, and antiplatelets (Mendelson and Prabhakaran, 2021), carry the risk of increasing the frequency and severity of HT (Moulin et al., 1994). Consequently, HT contributes significantly to the underutilization of reperfusion therapies and is associated with poor prognosis (Yaghi et al., 2015). Studies indicate that HT increases the risk of death by 8- to 10-fold (Yen et al., 2016), with evidence suggesting a higher bleeding propensity in Asian populations compared to Western populations (Shen et al., 2007; Ueshima and Matsuo, 2002; Kim et al., 2015). While symptomatic HT has garnered significant attention, the significance of asymptomatic HT remains controversial (Hacke et al., 2004; Berger et al., 2001; Albers et al., 2006; Molina et al., 2002; von Kummer, 2002; Dzialowski et al., 2007), despite its association with unfavorable long-term cognitive and neurological outcomes (Lei et al., 2014; Dzialowski et al., 2007).

Neuroimaging remains the principal method for predicting HT. Numerous studies have attempted to predict HT using non-contrast imaging signs, such as the hyperdense artery sign (Strbian et al., 2012), leukoaraiosis (Whiteley et al., 2012), collateral circulation (Bang et al., 2011), and hyperintense acute injury markers (Kidwell et al., 2008). Traditionally, logistic regression was the standard for analyzing prognostic data (Asadi et al., 2014). However, the emergence of machine learning and deep learning algorithms has demonstrated superior potential in outcome prediction and clinical decision support (Zihni et al., 2020; Khera et al., 2021). For instance, James et al. successfully employed support vector machines (SVM) to predict symptomatic HT following thrombolysis (James et al., 2018).

CT is an efficient, widely available diagnostic tool and serves as the primary imaging modality for acute ischemic stroke (Vilela and Rowley, 2017). In many resource-limited settings, non-contrast CT (NCCT) is often the sole imaging option available for initial evaluation. Given the variability in onset-to-visit times, we conducted a retrospective study using NCCT images obtained at various time points. Our study aims to present a novel deep-learning approach for predicting post-stroke HT using NCCT. This tool is intended to assist emergency neurologists in making informed medication decisions for high-risk patients.

Figure 1 describes this prediction system in detail. The prediction system consists convolutional neural network (CNN) and Residual blocks, which are used to describe the deep characteristics of CT.

Table 1 lists the details of the dataset. The cohort included 474 patients [mean age 62.45 ± 11.85 years; 123 (25.95%) females]. Treatment modalities included thrombolysis (101, 21.31%), antiplatelet therapy (314, 66.24%), and anticoagulation (12, 2.53%). The dataset contained 611 CT images acquired at various time points, comprising 243 patients in the HT group and 231 in the non-HT group. The training set included 437 patients (548 images), while the test set included 61 patients (61 images).

This study demonstrates the feasibility of using deep learning to predict HT after ischemic stroke, identifying potential risks even under conditions of complex and limited information. This tool aims to assist clinicians in individualized diagnosis, risk stratification, and the management of high-risk patients. Our model outperformed physicians in accuracy when utilizing only image information, presenting a novel approach for clinical HT prediction.

Unlike studies utilizing perfusion MRI (Yu et al., 2018), our approach employs NCCT, which is more accessible, faster, and widely applicable. While Dharmasaroja and Dharmasaroja (2012) achieved an AUC of 0.787 using machine learning on clinical data, a gap remains compared to CTNet's performance. Furthermore, while previous imaging studies focused on local features (Strbian et al., 2012; Whiteley et al., 2012; Bang et al., 2011; Kidwell et al., 2008), our deep learning approach analyzes complete CT images, allowing the network to learn a broader range of radiological features and their complex interconnections. Due to the unavailability of comprehensive clinical data, we were unable to assess clinical scale scores for our dataset. However, a multicenter trial of 3,035 AIS patients evaluated various scales for symptomatic HT, reporting AUCs ranging from 0.68 (IST-3) to 0.56 (SPAN-100) (Whiteley et al., 2014). In comparison, CTNet achieved an AUC of 0.842 (95% CI: 0.758–0.921), demonstrating superior predictive capability.

Using attention heatmaps, we visualized the anatomical regions utilized by the model. In cases with visible lesions (Figure 3a), the algorithm's focus correlated with the infarct area, confirming that the model effectively prioritized relevant pathological features. Notably, the model outperformed specialists in cases with low lesion visibility (Figure 3b), where it attended to specific brain regions that doctors might overlook. This suggests that deep learning attention maps could uncover subtle early indicators of HT.

Deep neural networks (DNNs) have demonstrated significant potential in predicting hemorrhagic transformation (HT) following ischemic stroke, owing to their capacity to extract intricate features from extensive medical imaging datasets. Nevertheless, the current model requires further training and optimization to meet the rigorous demands of clinical application. Specifically, future research should prioritize enhancing the model's accuracy and generalizability by integrating heterogeneous data sources and refining the network architecture. Furthermore, validating the model within real-world clinical workflows is imperative to ascertain its feasibility and effectiveness in supporting decision-making. Ultimately, the development of robust and precise deep learning frameworks for HT prediction holds the promise of improving stroke management through earlier detection and personalized treatment strategies, thereby optimizing both short- and long-term patient outcomes.

Despite exhibiting promising performance, our current work still has some limitations. Firstly, due to the limited sample size, we were unable to stratify patients according to their different treatment plans, which can have a significant impact on the occurrence of HT. Secondly, we did not perform segmentation of the lesion area, as we considered that unknown features outside the lesion area may also contribute to the prediction accuracy. In future studies, it may be worth investigating the use of lesion segmentation to further enhance the model's predictive performance. Thirdly, we have yet to establish a model built on a multi-center database, which may lead to limitations in the generalization ability of our model. Our future objective is to develop a model that can incorporate multi-modal data. This will require integrating a comprehensive and easily obtainable set of clinical data with imaging data to improve the accuracy and robustness of our model. Moreover, in analyzing the heatmap of attention, we observed that the model displayed attention toward areas corresponding to the skull, potentially affecting its performance. This indicates a need to remove the skull during image preprocessing. Ultimately, we aim to apply this model in a clinical setting to provide individualized treatment plan support and assist in decision-making for each patient.