Brain stroke is a sudden interruption of blood flow in the brain, leading to a lack of oxygen and nutrients to brain tissue. This can be caused by a blockage, which then is called ischemic stroke, or bleeding, which on the other hand is called hemorrhagic stroke (Alsbrook, 2023). Ischemic strokes are the most common type, accounting for approximately 88 % of all stroke cases, and they are the most known life-threatening neurological condition (Thiyagarajan and Murugan, 2021, Hui et al., 2024). Stroke typically evolves along multiple stages. The acute stage refers to effects occurring typically a few hours/days after stroke onset, while the sub-acute stage encompasses the period of two weeks to three months following stroke onset. In contrast, the chronic stage is characterized by long-term effects occurring months or years after the initial insult (Zhang et al., 2022). Other classification schemes use a much refined five-stage classification, including hyperacute (within 24 h after stroke onset), acute (within two weeks), early subacute (from two weeks to three months), late subacute (from three months to six months), and chronic stage (after six months). Fig. 1 presents a widely used classification of stroke types, also adopted here, with the categories highlighted in blue indicating the specific stroke types and stages that are the focus of this review.

Stroke patients typically undergo neuroimaging techniques to differentiate between ischemic and hemorrhagic strokes in addition to identifying the location and the size of the stroke, as accurate diagnosis is critical for determining the appropriate treatment. Magnetic resonance imaging (MRI) and computed tomography (CT) are commonly used for this purpose, each providing unique insights into brain health. MRI, known for its superior soft tissue contrast, is particularly useful when the diagnosis is uncertain or when more detailed imaging is required. It can offer more comprehensive information than CT, particularly regarding the location, timing, and underlying mechanisms of the stroke (Inamdar, 2021).

An MRI sequence, or MRI modality, refers to a specific configuration of pulse sequences and pulsed field gradients, leading to a distinct appearance in the generated image. Within MRI modalities, diffusion-weighted imaging (DWI) is especially valuable, because it can detect early changes in brain tissue caused by ischemia with a high contrast, helping to assess the full extent of the stroke’s impact (Zafari-Ghadim et al., 2024).

Brain stroke segmentation is the process of precisely identifying and segmenting the stroke-affected regions, called lesions, in neuroimaging data. Accurate stroke segmentation is crucial because timely and precise identification of stroke lesions directly impacts diagnosis, treatment planning, and patient outcomes. Early and accurate detection of stroke-affected areas allows clinicians to determine the type, severity, and location of the stroke, which is essential for deciding appropriate interventions such as thrombolysis or thrombectomy. Misidentification or delay in segmentation can lead to ineffective treatment, potentially resulting in severe neurological damage, disability, or even death. Thus, reliable stroke segmentation is critical in improving survival rates and reducing long-term complications (Bousselham et al., 2022).

To detect and segment diffusion abnormalities in ischemic acute strokes, historically, various traditional machine learning algorithms have been used, such as support vector machine and random forest (Castillo et al., 2021). Recently, remarkable progress in deep learning (DL) for brain lesion segmentation has been made, outperforming traditional methods. Convolutional neural networks (CNNs) have played a pivotal role in this advancement by efficiently learning spatial hierarchies of features from imaging data, which is crucial for accurate segmentation (Kamnitsas, 2017). U-Net, a notable DL model, has made a significant improvement in the area of segmentation by incorporating coarse features through skip connections (Ronneberger et al., 2015). Further U-Net variants where adopted later to enhance feature utilization and focus more on local features (Castillo et al., 2021).

Recently, there has been a surge of interest in introducing transformers to vision tasks, following the great success of transformers in natural language processing (NLP). Vision transformers (ViT) were proposed by Dosovitskiy (2020) for image classification, splitting an image into multiple linearly embedded patches and feeding them into a standard transformer with positional embeddings, leading to an impressive performance on ImageNet (Deng et al., 2009). In semantic segmentation, (Zheng, 2021) proposed SETR to demonstrate the feasibility of using transformers in this task, adopting ViT as a backbone and incorporating several CNN decoders to enlarge feature resolution.

Between 2020 and 2024, 10 reviews have been identified in this field (Thiyagarajan and Murugan, 2021, Zhang et al., 2022, Inamdar et al., 2021, Zafari-Ghadim et al., 2024, Castillo et al., 2021, Malik et al., 2024, Karthik et al., 2020, Werdiger et al., 2024, Abbasi et al., 2023, Parmilla and Muneeswari, 2023), addressing various aspects of acute-subacute ischemic stroke segmentation. Most of these reviews primarily focus on listing the key papers published in this area. In contrast, this review distinguishes itself by providing a comprehensive examination of all studies in the field of acute-subacute ischemic MRI segmentation, regardless of their publication venue impact. It also offers an extensive benchmarking of model performance and an in-depth analysis of factors influencing performance, including input data dimensionality, loss functions, preprocessing methods, and augmentation techniques across the reviewed studies.

Provide a summary of all the methods published in the field, and provide a benchmark for their performance and effectiveness.

Propose a comprehensive taxonomy that classifies the reviewed models based on their method and category.

Serve as a comprehensive reference for researchers by compiling all available datasets, MRI modalities, evaluation metrics, loss functions, preprocessing, and augmentation techniques used by the published articles.

Identify ongoing challenges in the field and highlight critical research gaps, and offer future directions to improve the effectiveness and accuracy of the deep learning models in this field.