Use of artificial intelligence in CT image evaluation in stroke patients – current options

Summary

Artificial intelligence and its rapid development represent one of the most important technological advances of the current decade. It affects almost all aspects of life, including medicine. Artificial intelligence is widely applied in neuroradiology, particularly in stroke diagnosis. The primary purpose of its application in this area is to accelerate the interpretation process, increase diagnostic accuracy, and help to select the treatment strategy. Clinicians involved in the initial management of a stroke patient should be familiar with the technical principles and possible use of artificial intelligence in neuroimaging, and they should know the strengths and weaknesses of the technology. This article briefly presents methods of artificial intelligence used in visual data processing. The main goal of the publication is to present particular automated analyses used in the interpretation of diagnostic information taken from CT images. CT is the primary choice in stroke diagnostics for most medical departments. The presented analyses are a calculation of the ASPECT score and detection of a hyperdense artery sign from non-contrast CT scans, identification of large vessel occlusion and collateral score evaluation from CTA, and creation of perfusion maps from CT perfusion.

Keywords:

deep learning – machine learning – ischemic stroke – large vessel occlusion – artifi cial intelligence

This is an unauthorised machine translation into English made using the DeepL Translate Pro translator. The editors do not guarantee that the content of the article corresponds fully to the original language version.

 

Introduction

Stroke is the second most common cause of mortality worldwide and the level of chronic disability associated with it is also high [1]. In Europe, it affects 1.1 million people annually and the economic costs associated with their care have been estimated at 45 trillion [1] euros per year [2]. Imaging modalities are a crucial part of the diagnostic process of CMP[2], their role is to detect, characterize and determine the prognosis of both acute ischemic and hemorrhagic CMP [1]. In most centres, CT scanning is used as the first choice modality, mainly because of its speed and wide availability [3]. Native CT scanning is able to differentiate subtypes of CMP and lesions mimicking CMP. In the case of ischemic CMP, CTA of the carotid and vertebro-basilar basins is also standard. Overall, 24-46% of ischemic CMPs are due to large artery occlusion [4], which predominantly occurs in the anterior circulation. Randomized controlled trials have demonstrated the safety and efficacy of early mechanical thrombectomy [5]. In particular, the DAWN and DEFUSE-3 trials are considered revolutionary, based on which [3]the American Heart Association/American Stroke Association recommendations were updated in 2018. According to them, the time window for performing thrombectomy in selected patients was extended from six to 24 h [4]. Revascularization endovascular therapy (EVL)[4] benefits patients with moderate to severe clinical deficits who have limited core ischemia and extensive penumbra detected on imaging [6]. Thus, the time window is replaced by a tissue window in the decision-making process [4]. Advanced neuroimaging methods -⁠ perfusion CT or MR (fluid attenuated inversion recovery [FLAIR], diffusion weighted imaging [diffusion weighted imaging;[5] DWI], perfusion weighted imaging [perfusion weighted imaging;[6] PWI] ] -⁠ are implemented to assess the core and penumbra [7]. Related to the increasing role of imaging in the diagnosis of CMP are advances in image processing and the development of artificial intelligence (AI) algorithms that are capable of automatically extracting diagnostic information. AI algorithms represent an important tool to support radiologists in speeding up the diagnosis of CMP and in determining the correct decision about the intervention under consideration in a shorter time [4]. The faster the diagnosis is made and appropriate treatment is initiated, the better the clinical effect for the patient [8].

 

Artificial intelligence

Artificial intelligence is defined as the ability of machines to mimic the cognitive functions of humans -⁠ learning and problem solving. AI can be understood as a set of programs and tools that make software "smart" to the extent that an independent observer thinks the output was produced by a human [9]. The field of AI is considered to be the most revolutionary area in the healthcare industry over the last decade, and diagnostic imaging has been the largest contributor to this development [10]. The fact is that the number of examinations in radiology practice has increased dramatically in recent years and the workload in this segment is likely to continue to increase in the near future [11]. Closely related to this is the boom in AI methods. As of September 2023, 237 medical devices using AI in radiology have been documented to be approved by the Food and Drug Administration (FDA) [12]. In particular, machine learning and deep learning [10] are showing breakthrough performance in image data analysis, providing an efficient way for fast imaging analysis [1]. Neuroradiology is one of the leading subspecialties in radiology in terms of the number and variety of AI applications [13]. Especially CMP medicine is well suited for their application because of the huge amount of data and the multidisciplinary approach to treatment. Brain imaging, which is[7] a key factor in CMP management and forms the basis for many clinical decisions, is an attractive subject for AI techniques [14], including tools for triage, quantification, tracking, and prediction [1]. In a May 2023 publication[8], Chandrabhatla et al. list up to 20 FDA-approved technologies used in the imaging diagnosis of CMP [15].

 

Machine Learning

Machine learning is a subset of AI using statistical approaches to allow machines to optimize outcome prediction after being exposed to data and trained to recognize patterns [1]. It is a field where computers learn from an accumulation of data without being specifically programmed [16]. Machine learning algorithms (e.g., linear regression, logistic regression, clustering, support vector machine method, random forest) evolve with increasing exposure to data; they do not work solely on the basis of rules, but improve with experience; they learn to answer specific questions by evaluating large amounts of input data [9,17]. In machine learning, the variables used as input data are generally referred to as features and are usually determined by the scientific team. Since the performance of machines is variable depending on the specified features, the selection and extraction of suitable features from the dataset is very important. In radiology, for image data, various image features such as size, localization, shape and density or signal intensity of the lesion can be used for machine learning. Machines are also able to distinguish and use other image features such as texture information -⁠ e.g., signal intensity gradient and skew -⁠ that are not discernible to the human eye. Machine learning is divided into supervised learning and unsupervised learning [17]. These methods can be distinguished based on whether they use human feedback [14]. What is common to both is that they are data-driven and the decision making process itself is done with minimal human intervention [18]. Supervised learning (with a teacher) uses a training dataset that is pre-labeled by a human to perform a given task [14]. When the program is exposed to sample data of the same type, it uses the characteristics of the training set to predict a specific outcome or goal [16]. The labeling process is laborious and time-consuming for the human in charge [14]. An example is a set of brain CT scans that the radiologist classifies into different groups (e.g., intracranial hemorrhage present/absent) [17]. In contrast, unsupervised (teacherless)[9] learning does not use human-defined answers, but instead attempts to independently identify naturally occurring hidden patterns or clusters in large datasets that are usually invisible to humans [14,16]. This includes, for example, clustering, an algorithm in which images are sorted into multiple groups based on a similarity metric with no a priori known driving momentum behind the separation process [14].

 

Deep learning

Deep learning is a machine learning method using a specific architecture, namely some form of neural network [17]. This branch of AI mimics the human brain by using numerous layers of artificial neural networks. These are composed of nodes and arranged in interconnected input, hidden and output layers. Deep learning is referred to as deep because it has multiple hidden layers [14,16], which represent interneurons [17]. These layers collect data from the inputs and provide an output that can gradually change as the system learns new features from the data [9]. Unlike other traditional machine learning methods requiring manual feature extraction from the inputs, deep learning techniques learn these features independently directly from the data, without the need for selection [19]. Deep learning algorithms provide particularly exceptional performance in image analysis, equaling or even surpassing human performance. Therefore, radiology is considered a natural domain for the use of deep learning [20], and deep learning is the basis of most AI tools for image interpretation [9]. The most popular and successful subset of deep learning in medical imaging is convolutional neural networks [1], inspired by the mammalian visual cortex [17]. Unlike traditional machine learning methods, convolutional neural networks have the ability to automatically identify patterns in complex image datasets, combining feature selection and classification into a single algorithm [1]. Hidden layer[10] convolutional neural networks use convolution and subsampling (spatial shrinkage) operations to decompose an image into features containing the most valuable information [16]. The key purpose of convolutional layers is to extract distinguishing features (e.g., edges, lines, shapes) from the input image information [18]. The nodes of convolutional neural networks are connected in a geometric structure, with each node connected to only a small portion of the input, which distinguishes them from conventional neural networks where each node is connected to every value from the input [14,18]. In image processing, the nodes of the input layer are arranged to form a convolution of a small portion of the image (the kernel), this kernel then moves around the image to form the output value, the information moves from simple to more advanced layers [14]. Finally, in a fully connected layer, representing a traditional neural network, high-level reasoning is applied and all the features from the image are combined, and the output layer provides predictions. This structure has applications within radiology in categorizing lesions or conditions from imaging modalities and also in deciding whether a particular pixel belongs to the background or target class [16,18].

 

Artificial intelligence in ischemic CMP imaging

AI techniques can be applied at all levels of stroke management -⁠ from prehospital care, including transport to the EVL centre[11], to radiodiagnosis, to choice of treatment strategy and subsequent rehabilitation care [21]. Within imaging, AI can improve the technical aspect of image acquisition. Machine learning approaches are able to improve the speed and quality of scanning, image reconstruction, reduction of artefacts[12] that arise during scanning and reconstruction or reduction of metal artefacts [22]. They also allow both radiation dose minimization and contrast agent dose reduction while maintaining optimal image quality [16]. Another benefit is the algorithms assisting in image interpretation. The basic imaging questions that need to be answered as quickly as possible in acute ischemic CMP are the exclusion of intracranial hemorrhage, the extent of acute ischemia, the presence of large vessel occlusion, and the extent of tissue at risk [4,23]. However, daily practice demonstrates several limiting factors in this process -⁠ evaluation and interpretation of imaging data are time consuming, radiologists have varying levels of experience and expertise in neuroimaging, and in the case of admission of a patient with CMP in community, peripheral hospitals with less experience in diagnosing CMP, in addition to less experience, truncated human resources are a significant problem[13] [23]. A limitation of qualitative assessment of CMP is also its subjectivity. The above-mentioned factors support the introduction of high-speed automated analyses, outperforming conventional methodologies and reducing time to treatment. AI tools can automatically produce quantitative measurements such as the Alberta Stroke Program Early CT Score (ASPECT) from a native CT scan, detect large vessel occlusion or determine collateral scores[14] on CTA, and process perfusion maps from a perfusion CT scan to assess potentially salvageable brain tissue [16]. Random forests are the most commonly used method to calculate ASPECT scores, and convolutional neural networks are most commonly used to detect large artery occlusion [24]. Especially since the discovery of deep learning, the evaluation of medical imaging using AI is considered a very rapidly expanding industry. Within the European Economic Area, before medical software can be marketed[15], it is necessary to obtain a certification mark that demonstrates compliance with medical device regulations approved by the European Parliament and the Council [8]. The most well-known commercial platforms used in the CMP diagnostic process include e-Stroke Suite (Brainomix, Oxford, England) in collaboration with Olea Sphere (Olea Medical Solutions, La Ciotat, France), Viz.ai (Viz.ai, SanFrancisco, CA, USA), RapidAI (iSchema View, MenloPark, CA, USA) [21]. These software are interfaced[16] with the hospital's Picture Archiving and Communication System (PACS), to which the assessment results are automatically transferred. They are available within minutes on work console screens and mobile applications. The AI tools are able to send a notification to the on-call neurointervention team, who in this way have the image documentation available for viewing virtually anywhere.[17] This step has also been shown to speed up the initiation of the therapeutic process [23]. In the study by Elijovich et al. the median time from CTA execution to sending the notification by AI was 19 min shorter than under conventional conditions (7 vs. 26 min; p < 0.001) [25].

 

ASPECT score

ASPECT is a quantitative scoring system implemented to more objectively [18]assess early ischemic changes in patients with arterio cerebral media (ACM)[19] occlusion on native CT scan [26,27]. In patients in the early time window (within 6 h of symptom onset), the ASPECT score is a crucial tool in patient selection for EVL based on several randomized clinical trials. According to current international guidelines, thrombectomy is indicated in patients with a score ≥ 6 [26-28]. Recently, the results of the multicenter randomized unblinded TENSION clinical trial have also been published, in which the beneficial effect of endovascular thrombectomy was demonstrated even in patients with proven large infarcts (ASPECT 3-5) and with a prolonged time window (up to 12 h) [29]. Despite conceptual simplicity, detection of early ischemic changes and determination of the ASPECT score is challenging in practice, especially for less experienced and trained evaluating physicians without a specialty[20] in neuroimaging. Unclear boundaries between the different regions included in the scoring, discrete early-stage densitometric changes in the ictus difficult to detect with the naked eye, time stress, assessor bias towards the expected findings, and differences in technical factors affecting image quality (X-ray energy, image processing, reconstruction algorithms) are among the main factors that cause inconsistency in scoring among radiologists [3,30]. The degree of agreement has been determined to be moderate[21] to moderate by several studies [26-28]. However, this may have clinical implications [28]. One way leading to a reduction of this variability and to an increase in the reliability of ASPECT score interpretation is the use of AI (Fig. 1), in particular deep learning techniques [3,27]. Available software programs for calculating ASPECT scores achieve promising results compared to human assessment. In a study by Nagel et al. the ability to detect early ischemic changes was compared between the commercial machine learning-based e-ASPECTS Brainomix system and three neuroradiologists, and no inferiority of automated assessment was noted [31]. In another study, automated assessment also achieved similar results to consensus reading by two experienced neuroradiologists [27]. In their publication, Maegerlein et al. even present a greater, almost complete agreement (k = 0.90) of the automated software with a predefined reference standard compared to two experienced neuroradiologists [28]. In contrast, Li et al. detected a lower success rate of automated software in a sample of 61 patients compared to two experienced radiologists [32]. In the work of Maegerlein et al. the relationship between the time interval of imaging from symptom onset and the ability to detect early ischemic changes was also investigated. When CT scans obtained within the first hour of symptom onset were analysed, both the software and the radiologist showed poor agreement with the consensus score. In the interval 1-4 h after symptom onset, AI analysis already showed high agreement with the consensus score (k = 0.78), whereas human performance was rated as poor to fair.[22] When imaging was performed after more than 4 h, the performance of both the analysis software and the radiologist was already comparable [28]. When comparing the different commercial platforms (Syngo.via Frontier ASPECT Score Prototype V2, Brainomix e-ASPECTS and RAPID ASPECTS, and Frontier ASPECTS Prototype and e-ASPECT Brainomix, respectively), the e-ASPECTS Brainomix software achieved the best results [33,34]. Automated assessment of ASPECT scores[23] may be less accurate in the presence of preexisting cerebral changes (in the case of leukoencephalopathy, in the field of changes after older infarcts or other tissue damage) or in the presence of spirals after coiling of an aneurysm [1,23]. It is also, like the radiologist's assessment, influenced by image quality, as confirmed by the results of a study assessing the effect of different reconstruction algorithms on the ASPECT score by four diagnosticians and the e-ASPECT Brainomix software [26]. The effect of slice thickness on software performance is also documented. As the slice thickness increases, its performance decreases significantly [35].

 

Detection of large artery occlusion

Large artery occlusion is the cause of approximately one-third of ischemic strokes. However, if it is not adequately treated in time, it causes severe neurological disability and accounts for up to 90% of ischaemic[24] stroke-associated mortality. A highly effective therapeutic procedure in these patients is endovascular thrombectomy, limited by a time window of no more than 24 h from the onset of symptoms. For these reasons, prompt and accurate diagnostic imaging is essential, especially at peripheral sites to ensure the fastest possible transport to an EVL[25]-enabled centre [36,37]. The detection of a large artery occlusion does not pose a diagnostic challenge for the expert in neuroimaging; the main benefit of AI technologies lies in the acceleration of the diagnostic and decision-making process. They also serve as a support tool for less experienced diagnosticians [8]. Large-caliber artery occlusion can be detected as early as on native CT imaging as a sign of a hyperdense vessel[26] (Fig. 2). The sign of a hyperdense vessel is a manifestation of erythrocyte-rich intra-arterial thromboembolus [38,39]. It is one of the early markers of ischaemic stroke, recognizable in contrast to ischaemic tissue changes immediately after arterial occlusion.[27] Thus, it represents a perfect diagnostic support in time-critical cases of ischemic CMP, especially in small peripheral hospitals where 24/7 availability of CTA is not assured [8]. It helps in faster identification of candidates for reperfusion therapy, without the need for administration of contrast agent and additional radiation exposure. Some platforms also provide information on thrombus length and volume, with thrombus volume and density being predictors of failed recanalization after thrombolysis administration, which predicts the need for interventional intervention[28] [40]. Commercial software for detection of hyperdense vessel symptom has reported sensitivity of 70-97.5% and specificity of 71-96% in studies; studies have also documented significantly shorter evaluation time of software compared to radiologists [38-40]. However, according to current recommendations, non-invasive angiographic imaging should confirm the presence of a blockage[29] [40], with CTA being the standard of care. Machine learning algorithms can also be applied in its evaluation (Figure 3) [8]. In recent years, several papers have been published assessing the reliability of several commercial software [30] in the detection of large artery occlusion in the anterior circulation (M1 segment of the arteria cerebri media [ACM][31]; the extent of ACM occlusion detected varies from study to study, some including[32] occlusion of the proximal M2 segment) from CTA, based on comparison with the assessment of experienced diagnosticians. The studies detected moderate to high sensitivity and specificity in detecting large artery occlusion (table 1); they agree on the lower ability to detect more peripheral occlusions, which are also potentially endovascularly treatable. The results of these papers support the use of AI programs as a complementary means to expedite diagnosis[33]. Currently, their diagnostic accuracy is not sufficient to replace expert evaluation by an experienced, board-certified neurodiagnostic specialist[34] [6,36-42]. Seker et al. report that the performance of the e-CTA Brainomix program is comparable to that of radiologists in specialty[35] training [42]. False-positive detected occlusions are mainly due to vessel asymmetry, which is easily distinguished by the radiologist by visual inspection from a large artery occlusion.[36] False negative findings can be explained by the inability of the algorithm to detect the occlusion when there is insufficient reduction in vessel density -⁠ in the case of incomplete occlusion or abundant collateral circulation, in the case of petrous and clinoid segment of the cerebri media[37], the false negativity may also be due to inadequate performance of the bone mask [36]. Future challenges for AI tools in this domain are mainly to optimize the detection of distal occlusions and occlusions in the vertebro-basilar basin, to differentiate occlusion from atherosclerosis, and to differentiate acute and chronic occlusion [22,40].

 

Evaluation of collateral circulation

One of the key factors on which the effect of thrombectomy depends is the status of collateral circulation [43]. Poor collateral circulation is associated with rapid progression of infarction; in this group of patients, urgent diagnosis and treatment are essential for tissue salvage. Conversely, abundant collaterals predict slow infarct progression, a longer time window,[38] and a better functional outcome of recanalization [36]. For optimal assessment of collateral circulation, scanning should be performed in the late arterial or early venous phase [8]. To assess the status of collaterals in the basin of the closed ACM[39] from CTA, a four-level scoring system is most commonly used, Tan score -⁠ from 0 (no collaterals present) to 3 (100% filling of the affected area)[40] [44]. This is a time-consuming and often difficult process due to the complex neurovasculature of the intracranial vasculature [45]. A particular problem with visual scoring of collateral circulation by radiologists is the inconsistency between raters, as it is a subjective assessment. One solution to this variability is to incorporate machine learning software into the scoring process (Figure 4) [46]. Objective, automated calculation of collateral scores has been investigated in several studies. In the work by Grunwald et al. [41] the degree of agreement between the software and a reference standard determined by consensus among three experienced neuroradiologists was assessed. The result was a 90% agreement rate of the software with the reference standard [46]. In another study, the performance of the software and 29 radiologists with different lengths of experience was compared against a reference standard determined by two, or in the case of equivocal agreement by three, independent neuroradiologists. The AI program performed similarly to the radiologists. An interesting finding of this study was that after 1 hour of training, there was no difference in assessment accuracy depending on the experience of the assessor [43]. Jabal et al. noted that the use of the software led to a significant increase in scoring accuracy and a reduction in inter-rater variability [47]. These findings support the implementation of AI methods in the assessment of collateral circulation[42] into clinical practice as a useful means[43] of reducing bias and identifying patients benefiting from thrombectomy [45].

 

Perfect[44] maps

In patients with an unknown time of onset of ischemic stroke or in patients in a prolonged time window (more than 6 h after symptom onset), cerebral perfusion imaging to assess the size of the nucleus and penumbra should be performed[45] as part of the selection process of suitable candidates before EVL[46] is considered [21,48]. MR imaging is considered to be the most accurate method. However, it is difficult to ensure its continuous availability in many departments. Therefore, a more accessible alternative, perfusion CT scanning, is routinely used in clinical practice [49]. Perfusion CT is based on dynamic tracking of the first passage of a bolus of contrast agent through the cerebral circulation. Several commercial software programs have been developed for the post-processing of perfusion CT data. The bulk of the currently used software[47] automatically generates[48] parametric maps (Fig. 5) and identifies the nucleus[49], penumbra, and their relationship to each other by deconvolution of tissue and arterial signals. Perfusion parameters in the affected area are compared with the contralateral hemisphere, and different applications use different quantitative thresholds to define the nucleus[50], producing different results even with the same source data, which can potentially influence treatment decisions [8]. For example, in the RAPID and Brainomix applications, an ischemic core is defined as tissue with at least a 70% reduction in blood flow (rCBF< 30%) compared to the unaffected hemisphere; Syngo.via CT Neuro Perfusion VB30 (Siemens Healthineers, Erlangen, Germany) uses an 80% reduction in flow (rCBF< 20%) as a threshold. All of the aforementioned software considers [51]hypoperfusion to be a prolongation of the time to maximal density of residual function -⁠ T max (the time for a bolus of contrast agent to pass from the proximal great artery to the cerebral tissue) to more than 6 s [50]. It must be emphasized[52] that most current techniques do not directly use AI algorithms. Their downside[53] is the sensitivity to noise, the need for human input in quality control of the contrast agent bolus saturation curves in the arterial and venous flow, and also the need to differentiate artifacts from the true perfusion deficit. Therefore, the current focus of research is to improve or replace current perfusion algorithms [48]. Among the already FDA-approved and CE (ConformitéEuropéenne) labeled software, Icobrain-CTP (Icometrix Leuven, Belgium) uses AI, specifically convolutional neural networks, to assess perfusion. The core and penumbral[54] volumes estimated by this software in the studies showed strong agreement with the results evaluated by radiologists [15]. An artificial neural network has been proposed by Kasasbeh et al. that is able to accurately predict the size of the core ischemia based on perfusion CT data and clinical input data (sex, age, National Institutes of Health Stroke Scale (NIHSS), time from symptom onset to imaging). It was tested in a sample of 128 patients and the mean absolute error between the core volume predicted by the neural network and the core volume detected from MR-DWI was 13.8 ml [51]. The use of deep learning algorithms to efficiently determine the nucleus[55] and penumbra on perfusion maps and their comparison with currently used methods was investigated by Bhurwani et al. The trained deep learning techniques were more accurate and outperformed the current methods [52]. Wouters et al. presented a deep neural network that is not only able to better predict the final infarct volume from perfusion CT compared to classical processing (compared to RAPID software), it can also predict the final infarct volume under different scenarios of recanalization success and different time interval to recanalization. Commonly used software can only predict final infarct size when complete recanalization is achieved or vice versa[56] in the absence of recanalization. The neural network created even generates models predicting the rate of infarct growth, the so-called 'tissue clock'. This feature may be helpful, for example, in deciding the need for repeat neuroimaging in patients transported to an EVL[57] center [47].

 

Hemorrhage

In patients with suppurative[58] stroke, an essential role of the initially performed native CT scan is to exclude the hemorrhagic form of ictus before thrombolytic therapy is administered. Even in the presence of intracranial hemorrhage, prompt interpretation of the findings[59] is one of the important factors influencing the clinical outcome of the patient [53,54]. To streamline the diagnostic process, software has [60]been developed capable of identifying hemorrhage from the native CT scan (Fig. 6), determining its volume, and alerting the evaluating physician to its presence [55]. Recent publications have highlighted their technical feasibility and their beneficial impact on worklist reprioritization, speed of assessment of findings, and length of hospital stay [56]. High accuracy of these algorithms in detecting hemorrhage has been reported in several datasets, with area under the curve (AUC) reaching up to 0.99, a sensitivity of 98%, and a specificity of 99% [54,57]. Studies cite[61] small hemorrhage volume (< 1.5 ml) and hemorrhage localized in the area of chronic pathological changes (older hematoma, area gliosis) as the cause of false-[62]negativity. False positivity has been reported in the case of calcifications, meningioma, hyperdense tumor mass, colloid cyst, aneurysm, thickened dura, imaging artifacts, or in the field of postoperative changes [55,58,59]. The algorithms achieved different levels of detection success depending on the type of hemorrhage -⁠ the highest, 100% sensitivity was demonstrated for the detection of intraventricular hemorrhage, while the detection of subarachnoid, subdural, and epidural hemorrhage was considered to be the most difficult [54,58]. Voter et al. observed reduced diagnostic performance of the software in patients with only solitary type of intracranial hemorrhage when compared with the occurrence of multiple types of hemorrhage in a single patient [59].

 

The future of artificial intelligence

The importance of AI in the diagnosis of CMP[63] is increasing. In particular, advances in critical decision making and patient choice of treatment strategy[64] are currently considered to be [65] challenges. For example, the development of a machine learning algorithm to identify patients who would benefit more from stent-retriever thrombectomy and those who are better candidates for aspiration, or an algorithm to recognize patients who have a higher risk of symptomatic intracerebral hemorrhage after thrombectomy, is desirable. Thus, the main goal of current AI perspectives is to optimize treatment -⁠ patient selection, prediction of outcome, and the ability to select an ideal group of patients who would not only meet the inclusion criteria for interventional procedure, but would have the least number and severity of complications [5]. The credibility of the use of AI products in clinical practice is limited by the fact that their scientific validation is more limited than approval by regulatory authorities [13]. Therefore, future clinical studies on large cohorts will be needed to validate and compare the performance of available software solutions, especially to extend and generalize the criteria for treatment selection in patients with stroke [8].

 

Conclusion

AI techniques should only be used in CMP diagnostics[66] in accordance with the approval of European and US authorities and thus only as a support tool for competent users. Their role is not to replace the radiologist; the final verdict on the diagnosis is always the decision of the evaluating physician [8,60]. They are beneficial both because of the rapid processing of large amounts of data, but they also increase the reliability of less experienced diagnosticians, which is particularly helpful in smaller, peripheral hospitals [26,36]. Incorporating AI tools into CMP[67] imaging diagnostics brings a reduction in inter-rater variability [15], increased efficiency, reduced time to treatment, and reduced errors, helping to improve the quality of patient care [16]. The disadvantage of this concept is that the evaluative skills of the radiologist depend heavily on the number of examinations evaluated and the accuracy of the visual image analysis. The results of some studies suggest that the use of AI software in the interpretation of findings reduces the vigilance of the evaluator. Thus, the novice radiologist may not acquire adequate interpretation skills and is at risk of becoming "addicted" to AI [8,9].

 

Conflict of interest

The authors declare that they have no potential conflict of interest.

 

Table 1. Results of studies investigating the reliability of commercial software in the detection of large artery occlusion in the anterior circulation from CTA.

Study	Software tested	Sensitivity	Specificity	Accuracy	PPV	NPV
Amukotuwa et al. [36]	RAPID CTA	94 %	765	-	43 %	98 %
Barreira et al. [37]	Viz-AI -⁠ Algorithm v3.04	90,1 %	82,5 %	86 %	81,8 %	90,6 %
Olive-Gadea et al. [40]	MethinksLVO, (Methinks Software S.L. Barcelona,Spain)	71,3 %	83,2 %	-	79 %	76 %
Rodriguez et al. [41]	Viz-LVO Algorithm® v1.4	87,6 %	88,5 %	87,9 %	-	-
Seker et al. [42]	eCTABrainomix	84 %	96 %	89 %	96 %	84 %
Yahav-Dovrat et al. [6]	See LVO	81 %	96 %	94 %	65 %	99 %