In the initial identification phase of the PRISMA flowchart (Figure 4), we identified 219 articles. During the screening phase, we reviewed the titles and abstracts of these publications to remove duplicates. We also excluded articles involving animals, other lesion types (eg, skin and lung), other stroke types (eg, perinatal stroke and chronic stroke), and other imaging modalities (eg, digital subtraction angiography). In addition, books, theses, review articles, and studies that did not propose an automated method based on ML or DL were excluded.

In the final inclusion phase, of the initial 219 articles, we retained 50 (22.8%) unique articles. These articles were thoroughly read to assess their eligibility for inclusion in this review. Articles were selected if they (1) relied on ML-based or DL-based automated methods and (2) included clinical records or medical images.

Stroke lesion segmentation is a challenging task due to the blurred boundaries between stroke lesions and normal tissues, particularly in hemorrhagic strokes. MRI images offer greater contrast and clarity compared to CT scans, allowing for more clearly delineated observation of stroke lesion boundaries [51].

Traditional thresholding and heuristic algorithms are inadequate for accurately segmenting stroke lesions. Before 2012, there was limited research on stroke lesion segmentation [52,53].

From 2012 to 2016, stroke lesion segmentation algorithms were primarily based on ML algorithms, with common segmentation methods mainly involving support vector machines (SVMs) [54], random forests (RFs) [55], and naive Bayes [56]. Due to the limited feature extraction capabilities of ML segmentation algorithms during this period, these algorithms primarily used MRI images.

In 2014, Maier et al [28] proposed an image segmentation method for ischemic stroke lesions based on SVMs using multimodal MRI images. The authors used SVMs to extract local features from the multimodal MRI data [28]. In the same year, Mitra et al [29] used a Bayesian-Markov random field for preliminary FLAIR image classification, then analyzed multimodal MRI data (T1WI, T2WI, FLAIR, and DWI) and contextually relevant features using an RF to identify likely lesion areas.

One year later, Maier et al [30] proposed another method for segmenting ischemic stroke lesions in multimodal MRI images using extra tree forests. This approach involved using extra trees for voxel-level classification and incorporated intensity-derived image features [30].

In 2016, Pustina et al [31] proposed an automatic segmentation algorithm for stroke lesions in T1WI images, combining RFs with multiresolution neighborhood data analysis to enhance efficiency and reduce observer dependency. In the same year, Griffis et al [32] used a stroke lesion segmentation method based on the Gaussian naive Bayes classifier. By using probabilistic tissue segmentation and image algebra to create feature maps, the authors encoded information about missing and abnormal tissues and used the leave-one-out method for cross-validation [32].

After 2016, there was a leap in the capabilities of DL for medical segmentation and feature extraction. DL has since outperformed ML as the main force in stroke image segmentation [57]. Due to the improved feature extraction capabilities of DL, algorithms for stroke lesion segmentation based on CT scans have emerged.

In 2018, Goncharov et al [33] proposed a CNN framework integrating UNet and T-Net (a variation of the UNet architecture). In each network, convolutional layers were replaced with residual blocks, and an initial convolution layer was added to enhance the learnability of the deep network [33].

In 2019, Clèrigues et al [34] proposed a 3D CTP segmentation method based on a 2D asymmetrical residual CNN. This method addresses issues such as small sample size, class imbalance, and overfitting by enhancing training, using symmetric modality augmentation, and applying uncertainty filtering, thereby improving performance and achieving fast inference [34].

In 2020, Wang et al [35] proposed a multimodal 3D CTP segmentation method based on CNNs. The authors extracted features from CTA images, synthesized pseudo-DWI, and used an innovative loss function to focus on lesion areas, thereby improving segmentation accuracy [35].

In 2021, Kuang et al [36] proposed a DL approach named EIS-Net, which combines a 3D triple CNN and a multiregion classification network for automatic segmentation of early ischemic stroke and Alberta Stroke Program Early CT Score scoring on NCCT in patients with acute ischemic stroke. In the same year, Soltanpour et al [37] proposed an improved version of the UNet network called MultiResUNet for the automatic segmentation of ischemic stroke lesions in CTP images. The algorithm uses multiscale convolutional blocks and incorporates contralateral CT images as references, estimating lesion locations using time-to-maximum heat maps [37]. Luo et al [38] proposed an ischemic stroke lesion segmentation method called UCATR for NCCT images. This method combines residual network (ResNet)-50 (a CNN that is 50 layers deep) and transformer encoders and introduces a multihead cross-attention module in the decoder to improve the accuracy of spatial information recovery [38].

In 2023, de Vries et al [39] proposed a model called Perf-UNet for segmenting the infarct core area from CTP source data. The model uses a symmetry-aware spatiotemporal convolutional structure, leveraging the dynamics of cerebral microperfusion, and incorporates attention modules in the skip connections [39].

In 2024, Kuang et al [40] proposed a hybrid CNN-transformer network based on circular feature interactions and bilateral difference learning. This network introduces a hybrid CNN-transformer in the encoder, a recurrent feature interaction module, and a shared CNN decoder with bilateral difference learning modules [40].

Compared to CT-based stroke lesion segmentation, MRI-based stroke lesion segmentation achieves higher segmentation accuracy.

In 2017, Chen et al [41] proposed a CNN network composed of a deconvolutional network for initial segmentation and a multiscale convolutional label evaluation network for eliminating false positives in small lesions.

In 2018, Zhang et al [42] proposed FC-DenseNet for segmenting acute stroke lesions in DWI medical images. This model combines dense connections and multiscale context on top of a 3D CNN to address common issues in DWI, such as noise, artifacts, and variations in lesion size and location [42].

In 2019, Karthik et al [43] proposed an improved fully convolutional network that applies leaky rectified linear unit activation in the last 2 layers of the UNet architecture. This approach precisely reconstructs ischemic lesions and allows the network to learn additional features not considered in the original UNet architecture [43].

In 2020, Clèrigues et al [44] addressed class imbalance issues by using symmetric modality enhancement, balanced training sample strategies, and dynamic weighted loss functions in the UNet architecture.

In 2023, Wu et al [45] introduced a lesion boundary–rendering method named TransRender. This method uses transformers to capture global information during the encoding phase and uses multiple renderings to effectively map encoding features of different levels to the original spatial resolution. The method adaptively selects points to compute boundary features based on point-wise rendering, supervising the rendering module to generate points and continuously refine uncertain regions [45]. In the same year, Wu et al [46] proposed a 2-phase brain multimodal MRI lesion segmentation method named W-Net. This method uses a CNN and a transformer as backbone networks, introduces a boundary deformation module and a boundary constraint module to address boundary ambiguity, and designs a multitask learning loss function to optimize W-Net from both regional and boundary perspectives [46]. Soh et al [47] proposed an algorithm named hybrid UNet transformer. This algorithm comprises 2 parallel pipelines, where the transformer-based pipeline uses feature maps from the intermediate layers of the UNet-based pipeline to enhance feature extraction capabilities [47].

The admission evaluation for patients with acute stroke typically includes stroke classification. Patients with acute ischemic stroke who are eligible for thrombolytic therapy should be referred to a thrombolytic center for treatment. For patients who are not suitable for thrombolytic therapy, conservative treatment should be provided upon admission. Patients with hemorrhagic stroke require rapid intervention to minimize brain damage and improve survival chances [24].

Acute ischemic stroke classification is typically based on CT imaging [58]. In CT images, ischemic stroke lesions appear as low-density regions (dark areas), while hemorrhagic stroke lesions appear as high-density regions (white areas; Figure 5) [48-50].

Adam et al [48] proposed a stroke lesion classification model based on decision tree and k-nearest neighbors (kNN) algorithms. The study found that the decision tree algorithm outperformed kNN in classification [48].

Gautam and Raman [49] proposed a CNN model for classifying stroke CT images. The authors improved the quality of CT images by using multifocus image fusion preprocessing and then inputted the processed images into a 13-layer CNN architecture for classification [49].

Chen et al [50] used hyperparameter optimization and transfer learning to identify stroke conditions in brain CT images. The optimized CNN and ResNet-50 models demonstrated high accuracy in distinguishing between normal, hemorrhagic stroke, ischemic stroke, and other lesions. Although ResNet-50 exhibited the highest accuracy, it required more processing time [50].

In clinical practice, the choice between using MRI and CT depends on the specific condition of the patient. CT has lower sensitivity in identifying early changes of acute ischemic stroke but demonstrates higher sensitivity in detecting characteristic high-density lesions of hemorrhagic stroke [59]. Compared to CT, MRI has higher contrast and a stronger ability to differentiate between brain tissues. The DWI sequence can detect minor changes in brain tissue just minutes after ischemia occurs. MRI also provides clearer identification of small hemorrhages and better differentiation between hemorrhagic and ischemic areas, along with detailed information on the extent and type of brain tissue damage after a stroke [60].

Before 2012, brain segmentation algorithms were primarily based on traditional methods for brain white matter segmentation. There are few studies focusing on the segmentation of stroke lesions in this period [52,53]. Between 2012 and 2016, segmentation algorithms for ischemic stroke lesions were primarily based on ML. Due to the blurred boundaries between stroke lesions and surrounding normal tissue in CT images, algorithms from this period mainly focused on MRI imaging. The accuracy of segmentation algorithms during this time generally did not exceed 0.75. These algorithms were complex, had weak generalization capabilities, and could not achieve the level of accuracy of manual segmentation. After 2016, with the development of UNet, DL has shown superior performance in the segmentation of stroke lesions [57]. As DL enables more in-depth feature extraction from medical images, significantly enhancing the extraction of effective features, an increasing number of researchers have started focusing on the segmentation of ischemic stroke lesions in CT images. However, the accuracy of these segmentation algorithms based on CT commonly did not surpass 0.75. Algorithms for segmenting ischemic stroke lesions based on MRI have shown better accuracy, enough to be comparable to manual segmentation. Some algorithms have achieved accuracies of up to 0.85. To facilitate the triage of patients with acute stroke, due to the urgency of the condition, the classification of stroke types primarily relies on CT images [58]. Current algorithms have achieved high accuracies, generally exceeding 0.90, indicating promising application prospects.

Data preprocessing is important for medical image segmentation. Preprocessing aims to standardize image quality, enhance valuable features, and reduce segmentation difficulties. Clèrigues et al [34] and Clèrigues et al [44] used symmetry-enhanced modalities, leveraging the high bilateral symmetry exhibited by the human brain in its natural state. This approach provides clearer and more accurate information for subsequent steps such as feature extraction, lesion detection, and image segmentation. For complex segmentation tasks such as stroke lesion segmentation, the introduction of multimodal approaches can significantly enhance algorithm performance by extracting complementary information from different imaging techniques. The studies by Mitra et al [29] and Wang et al [35] are particularly representative. Mitra et al [29] used multimodal MRI data, including T1WI, T2WI, FLAIR, and DWI, along with context-aware features. Wang et al [35] extracted features from CTA images, synthesized pseudo-DWI, for use in CTP segmentation methods.

In addition, some studies used multistage algorithms, combining coarse and fine-grained approaches to achieve better segmentation performance; for instance, Mitra et al [29] used a Bayesian-Markov random field model for preliminary classification of FLAIR images and used RFs for multimodal MRI segmentation. Chen et al [41] used a deconvolutional network for initial segmentation and used a multiscale convolutional label evaluation network to evaluate and eliminate false positives of small lesions, achieving promising results on large data sets and demonstrating the algorithm’s generalization ability.

Since 2021, various attention mechanisms and transformers have been widely applied in the field of medical image segmentation, leading to further improvements in the performance of stroke lesion segmentation algorithms. Transformers leverage self-attention mechanisms, enabling neural networks to capture global information and exhibit excellent performance when processing long sequence data [61]. De Vries et al [39] introduced attention mechanisms on the skip connections between the encoder and decoder, reducing the time dimension and only propagating the most informative features. Luo et al [38], Kuang et al [40], Wu et al [45], Wu et al [46], and Soh et al [47] all incorporated the concept of transformers, leading to significant improvements in algorithm performance.

Clinical methods for stroke risk prediction include clinical risk scores (eg, Framingham Stroke Risk Profile [62]), genetic risk analysis [63], serum biomarker detection [64], and imaging analysis [65]. Compared to other methods, imaging analysis provides direct anatomical and functional information, allowing clinicians to observe pathological conditions directly. In addition, imaging data can quantify stroke risk factors such as atrial fibrillation (AF), carotid artery stenosis (CAS), and cerebral small vessel disease, which are crucial for stroke risk prediction [66]. Furthermore, imaging techniques can monitor disease progression and treatment effects in real time, enabling clinicians to adjust treatment strategies promptly. Due to these advantages of imaging techniques, AI-based stroke risk prediction using medical imaging has become a major focus of research (Table 3).

Stroke can be divided into 3 distinct subtypes with well-defined etiologies: large artery atherosclerosis, cardioembolism, and small artery occlusion (Figure 6) [66].

Large artery atherosclerosis-type stroke is typically caused by atherosclerotic stenosis in the carotid and intracranial large arteries. For patients with mild stenosis, conservative treatment is recommended. For those with severe stenosis, carotid endarterectomy surgery is performed to prevent the occurrence of ischemic stroke [71].

Cardioembolism-type stroke is caused by cardiogenic diseases, where emboli from the heart and aortic arch dislodge and circulate, leading to cerebral artery embolism [83]. Studies have shown that left atrium (LA) enlargement may lead to myocardial fibrosis and disruption of atrial muscle bundles, resulting in blood stasis within the atrium and reduced atrial contraction capacity. This can lead to AF, thereby increasing the risk of cardioembolic stroke [84].

Small artery occlusion-type stroke, also known as lacunar stroke, is a pathological condition caused by disease in the small penetrating arteries of the deep regions of the cerebral hemispheres or brainstem, leading to ischemic necrosis of the brain tissue supplied by these arteries [85]. The international Standards for Reporting Vascular Changes on Neuroimaging-1 clearly state that white matter hyperintensities (WMH) are neuroimaging indicators of small vessel disease and can predict the occurrence of stroke [16].

The most commonly used imaging method for the detection of CAS in clinical practice is ultrasound (Figure 7) [86,87].

In 2015, Alam et al [67] proposed a robust fuzzy radial basis function network that combines fuzzy c-means clustering and a radial basis function network, incorporating spatial information and a smoothing parameter to handle noise.

In 2017, Araki et al [68] combined improved spatial domain filtering techniques with ML to use the morphological differences in carotid artery walls from ultrasound images. The authors used the improved spatial domain filtering techniques to remove image noise and used SVMs for classification [68].

In 2021, Al-Mohannadi et al [69] introduced morphological operations in an encoder-decoder structure to reduce the impact of noise on the results, using Sobel gradient direction images and Prewitt gradient direction images as inputs. In the same year, Mi et al [70] designed an MBFF-Net (a multibranch feature fusion network) with 3 branches. The first 2 branches extract plaque features at multiple scales and different contexts, while the third branch uses prior information [70].

In 2022, Latha et al [88] compared various ML and DL algorithms for carotid artery segmentation. Among the ML algorithms, RF performed the best, while CapsuleNet outperformed other DL algorithms [88].

In 2023, Jiang and Chiu [72] integrated a centerline extraction network and a dual-stream centerline-guided network into a 3D UNet. The centerline extraction network generates centerline heat maps to indicate the position of the carotid artery centerline, while the centerline-guided network segments the 3D ultrasound images based on the centerline heat maps [72].

Similar to CAS, echocardiography is commonly used clinically to detect stroke-related heart disease, such as AF and LA enlargement (Figure 8) [15].

In 2018, Degel et al [73] proposed an algorithm that combines DL, shape priors, and adversarial learning to achieve automatic segmentation of the LA in 3D ultrasound images. This algorithm uses VNet for 3D volume segmentation and trains an autoencoder network to constrain the segmentation results to the shape of the LA [73].

In 2019, Moradi et al [74] proposed MFP-UNet, which combines a pyramid network with the UNet model, adding 2 downsampling layers to extract more detailed features from the images. The authors also used the Niblack global thresholding method to preprocess the cardiac ultrasound images to enhance contrast [74].

In 2020, Leclerc et al [75] proposed an RUNet composed of 2 UNets. The first UNet obtains the region of interest by dilating the preliminary results, and the second UNet is used to predict the final segmentation. RUNet outperformed other networks in reducing geometric and anatomical prediction anomalies [75].

In 2021, Liu et al [76] introduced a pyramid-shaped local attention mechanism in UNet to capture contextual information and designed a label consistency learning mechanism to improve the classification accuracy of edge pixels in cardiac ultrasound images.

In 2023, Li et al [77] proposed a multitask DL model for cardiac ultrasound segmentation and key point detection, named EchoEFNet. The network introduces ResNets, branched networks, and dilated convolutions in the encoder part and incorporates multiscale concepts in the decoder part, combining both low-level and high-level features [77].

MRI is superior to CT in terms of soft tissue contrast, allowing for better differentiation between white matter and gray matter in the brain. Therefore, MRI is more commonly used clinically to detect cerebral small vessel disease and WMH (Figure 9) [89,90].

In 2020, Liu et al [78] proposed the M2DCNN, a network composed of 2 symmetrical U-shaped subnetworks. This network uses dense blocks, dilated blocks, and multiscale processing to improve feature extraction and prediction capabilities. In addition, it introduces a loss function based on the similarity between lesions and background [78].

In 2021, Park et al [79] proposed a UNet with multiscale foreground highlighting, incorporating an auxiliary classifier in the middle layers of the decoder for deep supervision training. The authors used multiscale label images and applied the highlight foreground method to enhance foreground voxels [79].

In 2021, Li et al [80] introduced squeeze-and-excitation blocks and dense connections into the UNet architecture to better capture spatial and multiscale semantic information.

In 2023, Huang et al [81] proposed the level set loss (LSLoss) loss function. LSLoss consists of 4 loss terms, each designed to optimize foreground loss, background loss, region of interest loss, and divergence loss, to improve the accuracy of segmentation results [81].

In 2024, Farkhani et al [82] proposed the volumetric segmentation of WMH using transformer architecture, which enhances the UNet by using consecutive self-attention encoders to capture spatial dependencies in the data.

Carotid and cardiac ultrasound segmentation can assist physicians in quantitatively assessing the condition of a patient’s heart and carotid arteries, providing important references for evaluating the risk of stroke. However, ultrasound image segmentation is a complex task. Compared to CT and MRI images, ultrasound images have lower contrast, noisy and blurry boundaries, and numerous artifacts, making ultrasound medical image segmentation more challenging.

For ultrasound image segmentation and classification, DL generally outperforms ML. Among ML algorithms, RF, with its capability to handle multifeature data, performs the best in handling these tasks with high accuracy and robustness. By contrast, other ML algorithms such as SVM and kNN perform well in certain specific tasks but are generally not as stable and accurate as RF. This is similar to the excellent performance of RF in stroke lesion segmentation and classification as discussed in the Segmentation and Classification of Stroke Lesions subsection, further demonstrating RF’s superior performance in image segmentation and classification tasks [88].

To address the significant noise in ultrasound images, many researchers have incorporated denoising techniques into their algorithms. Alam et al [67] introduced spatial information and smoothing parameters to handle noise, and Araki et al [68] used spatial filtering techniques. In addition, due to the high noise and blurry boundaries in ultrasound images, advanced architectures such as transformers may not necessarily outperform traditional convolutional blocks. In the comparative experiments in the study by Liu et al [76], traditional architectures such as UNet, UNet++, and ResNet performed better than Swin-UNet. Although Swin-UNet, with its advanced transformer architecture, significantly improved feature extraction and generalization capabilities, it exhibited overfitting in the segmentation of edge regions due to the substantial noise and blurred edges in ultrasound images.

The structures of carotid and cardiac ultrasound images are relatively fixed, and the introduction of morphological operations can better optimize these fixed structures. This effectively reduces the impact of noise and small artifacts, improves segmentation accuracy, enhances structural features, and increases robustness. Araki et al [68] used the morphological characteristics of the carotid artery wall to specifically improve spatial filtering, while Al-Mohannadi et al [69] introduced Sobel and Prewitt operators to make the model more attentive to edge information, aiding in the identification and segmentation of the boundaries of the carotid artery and the heart. Mi et al [70] and Degel et al [73] incorporated shape prior information, and Liu et al [76] introduced a label consistency learning mechanism. These 3 studies used the concept of templates to constrain the segmentation results and reduce missegmentation. For the segmentation of carotid vessels, centerline extraction can accurately determine the center position of the vessels, providing precise references for subsequent segmentation and feature extraction. This can largely ignore small noise and artifacts in the image, providing more stable segmentation results [72].

For WMH segmentation, many researchers have attempted to use attention mechanisms to enhance the model’s focus on small targets. The attention mechanism allows the model to selectively focus on important regions of the image based on the task requirements, reducing dependence on irrelevant information [78]. Farkhani et al [82] achieved the precise capture of key regions through self-attention and hybrid transformer mechanisms, contributing to the improvement of WMH segmentation accuracy.

In the task of WMH segmentation, there is a significant class imbalance due to the large difference in the number of pixels between white matter and non–white matter regions. The WMH segmentation model tends to predict the more numerous classes while ignoring the minority classes, leading to overfitting. To address this issue, Liu et al [78] and Li et al [80] introduced dense blocks, which reduce the number of training parameters and alleviate the problem of vanishing gradients.

For small target segmentation, optimizing the loss function is a promising approach. In the study by Huang et al [81], an LSLoss function was proposed to train the segmentation network. This method allows for training without actual ground truth labels, helping to avoid overfitting the training data.

Stroke prognosis refers to predicting long-term outcomes such as recovery status, quality of life, and survival rates of patients based on medical imaging and clinical dates. AI technology can quantify the extent of brain tissue damage, assess patients’ prognostic situations, and provide benchmarks for long-term rehabilitation (Table 4) [91].

Acute stroke lesions tend to increase in size over time, leading to increased damage to brain tissue. To address this issue, some researchers have used AI to predict the final infarct lesions and the penumbra [103].

In 2017, McKinley et al [92] proposed an RF-based algorithm for penumbra prediction in multimodal MRI, called FASTER. This algorithm first extracts various statistical metrics from perfusion and diffusion imaging in the MRI data and then uses a decision forest algorithm to predict the penumbra [92].

In 2018, Nielsen et al [93] used CNNdeep to predict final infarct lesion, outperforming traditional general linear models, CNNs based on time to maximum, and apparent diffusion coefficient threshold methods.

In 2020, Yu et al [94] proposed a UNet-based model for predicting the final infarct lesions in patients with acute ischemic stroke. This model is the first DL-based stroke infarct prediction model that does not rely on reperfusion information. The volume error between the predicted results and the ground truth is only 9 mm³ [94].

In 2021, Pinto et al [95] proposed a method combining restricted Boltzmann machine and gated recurrent unit to predict the final stroke lesion in patients with acute ischemic stroke after 90 days. Each restricted Boltzmann machine learns features from different MRI parameter maps, which are then merged with the original maps and input into a convolutional and recurrent neural network architecture for supervised learning [95].

In 2022, Wong et al [96] developed a DL-based method for predicting the final infarct volume from MRI images. This approach uses networks with rotation and reflection variants and incorporates clinical variables to predict the 90-day modified Rankin Scale score [96].

Clinical data can assist in estimating stroke recovery. Lesion volume and lesion location are critical factors in predicting recovery. Patients with extensive brain damage or strokes affecting critical brain areas generally have a poorer prognosis. Integrating patient clinical records with medical imaging can enhance the estimation of recovery outcomes [104].

Bentley et al [97] used SVMs to predict whether patients with acute ischemic stroke would develop symptomatic intracerebral hemorrhage after thrombolysis. Through the analysis of clinical records and brain CT images of 116 patients with acute stroke, the authors found that SVMs outperformed traditional prediction scoring systems in predicting symptomatic intracerebral hemorrhage [97].

Monteiro et al [98] investigated the use of ML methods to predict functional recovery in patients with ischemic stroke 3 months after the event. The study examined the performance of 5 different classifiers, among which RF exhibited the best performance [98].

Cheon et al [99] constructed a deep neural network model combined with principal component analysis for feature extraction. The model demonstrated high predictive ability based on clinical records from 15,099 patients with stroke [99].

Kuang et al [100] proposed an automated method for Alberta Stroke Program Early CT Score scoring from NCCT images of patients with acute ischemic stroke. This method uses texture features to train an RF classifier and performs well in identifying large areas of early infarction [100].

Scrutinio et al [101] used an RF algorithm enhanced with synthetic minority oversampling technique to predict the 3-year mortality rate of patients with severe stroke, significantly outperforming the logistic regression model.

Brugnara et al [102] integrated clinical records, multimodal imaging, and angiographic features to predict the modified Rankin Scale outcomes of patients with acute ischemic stroke after endovascular treatment. With the increase in medical information modalities, the model performance gradually improves [102].

For stroke prognosis, both DL and ML have their respective advantages [105]. DL models require large amounts of data for training to learn sufficiently complex features and avoid overfitting. This also explains why the deep neural network in the study by Cheon et al [99] achieved better performance after being trained on 15,099 clinical records, although acquiring a large amount of high-quality clinical data for stroke prognosis may be challenging.

By contrast, traditional ML methods have demonstrated greater robustness and efficiency in handling stroke prognosis tasks. Studies have shown that when handling stroke prognosis tasks, RF and its variant models have shown better performance than other ML algorithms [98,100,101]. This is mainly due to the robustness of RF models against data noise.

For the estimation of the final infarct lesion and penumbra, although the existing estimation may not achieve high DSC values, metrics such as volume error and AUC have reached high levels, indicating good practical value. In the future, through continual optimization and improvement of algorithms, the clinical translation prospects of AI in final infarct estimation will be even broader.