Computer-aided imaging analysis in acute ischemic stroke – background and clinical applications

CAD is strongly related to Artificial intelligence (AI), a branch of computer science that has witnessed an incredible development in the last few years. CAD and automated computer diagnosis (ACD) are two concepts with similar names but different meanings.

CAD usually relies on a combination of interpretation of medical images through computational algorithms and the physicians’ evaluation of the medical images. In this case, medical physicians are not replaced by an algorithm, but they use algorithms’ output as a second opinion. The diagnosis and the final decision are made at the end of the process by the physicians [3]. This concept is useful particularly in cases where physicians are less confident about the diagnosis so that the final decision may be improved by the use of algorithms’ results. However, it is worth noting that this approach can have downsides as well, and some studies suggest that users tend to become less vigilant when aware that of the CAD results while interpreting medical images [4].

ACD is based on computer algorithms only, and the diagnosis is made directly by the algorithm. In this case, the algorithm’s performance in a clinical routine must be at least equal or better than the performance of an average physician. Although computer algorithms may easily exceed human performance in many fields, developing such tools for the usage in medical imaging remains a difficult task. More difficult is even the adequate assessment and validation of CAD or ACD based algorithms.

In 2011 Goldenberg et al. presented another concept of CAD for the usage in emergency medicine called computer-aided simple triage (CAST), which performs an analysis of medical images and sorts them into different prioritisation’s categories. CAST systems should attract the radiologists’ attention to acute and time-sensitive critical cases [5].

Physicians use CAD systems as a support for their decision-making process aiming to get better results in the detection and interpretation of pathologies in medical images. Evaluating the performance of CAD schemes against a gold standard is usually not sufficient; measuring the influence of CAD systems on the decisions made by the physicians and on the general workflow in a clinical setting is also crucial. This influence may be positive, and this could be mirrored by a decrease of time need to diagnosis establishment or an increase in the detection rate of true subtle anomalies due to CAD support. A negative influence may be marked by an increase in false positive results due to the confusion generated by the bias effect. Therefore, although not easy, it is essential to study the effect of CAD systems on their users. For such studies, a large number of physicians or radiologists is needed, and a comparison between final results obtained by the physicians with or without getting CAD support in different settings is helpful to determine the impact of these tools on the performance of the physicians, which is related to the outcome of patients.

CAD algorithms are developed to perform tasks that usually require human intelligence and aim at extracting patterns from medical images and using these patterns to perform a specific task like suggesting a diagnosis.

Pattern recognition and extraction requires prior identification of relevant image features. Classical multistep CAD systems are based on conventional machine learning algorithms in which human-experts hand-engineer these features. Classical CADs process data in multi-steps (at least two: hand-encoded features extraction and classification) [6].

Unlike these, modern CAD systems use representation learning (RL) based algorithms, in which no manual feature encoding is necessary [7]. RL schemes determine the best features to use while classifying the input data on their own. Processing steps of modern CAD are sometimes not distinguishable, because of their structural properties usually based on neuronal networks with multiple hidden layers.

Deep learning (DL) is a subfield of RL, in which algorithms get from simple features, like edges or textures, to more complex features such as shapes or organs in their learning process (Fig. 1a and b) [8, 9]. Through deep learning astonishing results have been made possible in previously very challenging visual tasks such as the ImageNet challenge. Consequently, these methods have already been successfully applied to the medical field, e.g. for the detection of melanoma [10] and the detection of intracranial haemorrhage [7].

Classical and modern CAD systems alike are usually trained by using labelled data; this method is called supervised learning. The labelling is commonly done by a human expert; the algorithm’s possible output-results in this model are well defined. It is worth noting that this is a crucial step in developing e.g. deep learning-based algorithms. It can generally be said, that a larger dataset for training should lead to a more robust algorithm that would be less prone to overfitting and should be able to perform better on previously unseen external data. However, large datasets are not easy to obtain, and labelling large datasets is time-consuming if done by hand, but prone to errors if done automatically. On the other hand, unsupervised learning algorithms try to discover previously unknown patterns and structures in the unlabelled input data without previous labelling. The algorithm decides by itself how to cluster the data into different subgroups. A combination of these two methods is called semi-supervised learning, in which a large amount of unlabelled data in conjunction with a usually small quantity of labelled data are used [9]; this method could merge benefits of both previously cited approaches (more accuracy as in supervised learning, and less time for data labelling as in unsupervised learning).

Pattern recognition process in classical and modern CAD medical algorithms usually follows three main steps; however, getting through all these steps is not mandatory. These main steps are: (i) preprocessing of medical images – including segmentation and designation of regions of interest (ROI), (ii) extracting automatically generated or hand-engineered features that are predefined from human experts and finally (iii) data classification based on these features (Fig. 2). Modern CAD algorithms can present output data without necessarily getting through all these steps; this was made possible after the introduction of neural networks with multiple hidden layers.

Especially since the advent of deep learning, medical image analysis using AI is a worldwide rapidly growing market [11]. Several software tools have been made commercially available that aim to support radiologists and medical doctors in making more rapid and precise decisions in the diagnosis of stroke, which could be beneficial for patients’ outcome. Table 1 lists main companies and their presented applications.

Many methods have been applicated to evaluate the performance of CAD systems for commercialization. Some examples of these methods are leave-one-out, cross-validation, hold-out, and resubstitution. However there is until now no standardized approach; trying to solve this issue, the Computer Aided Detection in Diagnostic Imaging Subcommittee (CADSC) which is a committee initiated by the American Association of Physicists in Medicine (AAPM) proposed some recommendations on the methodology applicated in the evaluation of CAD system performance.

Health authorities firmly regulate the commercialisation of medical devices and drugs. Medical softwares are usually included under the medical devices category. Software as a Medical Device (SaMD) is a new term defined by the International Medical Device Regulators Forum (IMDRF) for applications that are used without being a part of a hardware medical device. These SaMD have special regulations and validation processes, which are adopted by the American Food and Drug Administration (FDA). Within the European Economic Area, medical softwares must obtain a certification mark (CE Marking) demonstrating conformity with medical devices regulations (MDR) approved by the European Parliament and Council before free commercialization. Of note, some of these certification processes do not require the presentation of clinical validation data.

A non-contrast-enhanced computed tomography (NCCT) brain scan is still the most widely available tool in acute stroke imaging because it is easily accessible, inexpensive, efficient, fast and reliably rules out haemorrhage. The most accurate assessment of the early infarction is obtained by diffusion-weighted magnetic resonance imaging (DWI) [12]. However, DWI is not everywhere available in the acute setting. Quantitative measurements of acute infarct on NCCT are difficult in clinical routine, since signs of infarction are more subtle and human assessment is highly variable. Hence, the correct NCCT interpretation of a patient with an acute ischemic stroke before thrombolysis or thrombectomy requires training and experience.

The probably first semi-automated approach to identify putative hypodensity within the middle cerebral artery (MCA) territory was published in 2001. In the following years, several other different computer-aided detection schemes for cerebral ischemia on CT were published. However, all of these papers described different approaches and focused on the methodology of the algorithms used; sample sizes were rather small and rigorous comparisons against the current gold standard, the interpretation of the scan by a neuroradiologist or against DWI were missing (Additional file 1). Up to date, only two commercial products are available that are certified for use in clinical routine: the e-ASPECTS® software from Brainomix Ltd. (Oxford, UK) and RAPID ASPECTS® by iSchemaView (Menlo Park, USA).

Siemens developed another post-processing tool for early ischemic change detection in CT using the ASPECT score (syngo.via Frontier ASPECT Score Prototype V1_2_0, Siemens Healthcare GmbH, Erlangen, Germany), which is not yet certified for clinical application but has undergone comparison to the e-ASPECTS software [13]. Here the authors found high agreement in ASPECTS rating between two certified radiologists, expert consensus reading of NCCT images, and e-ASPECTS, but only low to moderate agreement to Frontier-ASPECTS by Siemens.

e-ASPECTS, RAPID ASPECTS, as well as Frontier-ASPECTS, are based on quantitative evaluation of early focal ischemic damage by the Alberta Stroke Programme Early CT Score (ASPECTS), which is a topographic scoring system that divides the MCA territory into ten areas of interest. Originally ASPECTS was calculated within two prespecified slices through the level of the basal ganglia and the level of supra-ganglionic structures [14], while softwares are now integrating the whole brain scan and visually highlight the damaged ASPECTS region. e-ASPECTS from version 7 also displays acute ischemic volume in millilitres illustrated by a coloured heat map. The automated assessment overcomes the significant intra- and interrater variability of ASPECTS and hence standardizes the clinical application [15, 16]. Current guidelines recommend ASPECTS as an imaging selection criterion for mechanical thrombectomy (MT) in patients within 6 h from stroke onset [17].

While so far only one study on the performance of RAPID ASPECTS has been published [18], there are several studies on the performance of e-ASPECTS within different settings and patient populations available [19‐24]. All these studies indicate that these algorithms can be better than non-stroke experts and at least equal than experts in applying the ASPECTS to patients with acute ischemic stroke, yet they are not intended as a stand-alone diagnostic tool. Furthermore, e-ASPECTS has been compared to CT perfusion with regard to prediction of clinical outcome and final infarct size in patients with large vessel occlusion undergoing MT [25, 26]. The results of both studies suggest that NCCT based infarct cores estimation can be an alternative to computed tomography perfusion (CTP) derived infarct core estimation. Recently it was also shown that e-ASPECTS ratings and further clinical criteria could be successfully used to identify suitable candidates for MT in patients with longer or unknown time windows [27]. Importantly, the only studies showing an increase of physician’s performance before and after the aid of an automated algorithm are available for e-ASPECTS [28] and e-CTA® [29].

Hyperdense vessel sign (HDVS) on NCCT represents an early marker of acute ischemic stroke caused by intracranial arterial occlusion. HDVS is a radiological phenomenon marked by an increase of vessel radiodensity on NCCT after an acute occlusion. HDVS can be seen in various vascular diseases, including acute arterial occlusion, acute arterial dissection, aneurysm rupture, and acute venous thrombosis [30].

HDVS is most commonly reported in the MCA region; this is because MCA territory is usually the most affected cerebral region by ischemic stroke [31]. Also, the MCA has a large diameter in comparison to other intracerebral arteries and the majority of its branches run parallel to the –commonly most reconstructed– transverse imaging plan in cranial CT.

HDVS is highly specific (95%) and moderately sensitive (55%) for arterial obstruction in acute ischemic stroke; usage of thin-slicing improves the sensitivity significantly [32]. However, physiological calcifications or hyperdense structures outside cranial arteries are frequent. Koo et al. defined some objective criteria of MCA HDVS to differentiate it from normal MCA vessels. These are: In NCCT, the density of the pathological MCA should be superior to 43 Hounsfield units (HU) and 1.2 times higher than the contralateral MCA [33]. Lim et al. published a work about the value of HDVS in detecting large vessel occlusions (LVO) in the setting of neurological acute ischemic presentation, especially for hospitals with no access to CT-Angiography. They concluded that the HDVS has a high sensitivity and specificity for recognizing LVO on thin-slice NCCT in acute ischemic stroke patients presenting with an NIHSS more than 10 and suspected occlusion of MCA (M1 segment) or basilar artery [34].

Through its early visibility, already after vessel occlusion and before upcoming pathological parenchymal ischemic changes, it represents a perfect diagnostic aid in time crucial acute stroke cases. Automating the process of MCA or intracranial HDVS sign detection in emergency imaging may accelerate the identification of positive cases, especially in spoke centers without regular access to CTA. This could improve referral logistics and reduce the time to acute treatment, e. g. mechanical recanalization or systemic thrombolysis. Important success elements of CAD of HDVS are: Acquisition of high-quality cranial native CT, application of new reconstruction methods like iterative model reconstruction to reduce noise and improve diagnostic performance [35] and using thin slicing to improve the sensitivity of HDVS. Studies suggest that potential algorithms can achieve a sensitivity if up to 97,5% in detecting HDVS (Additional file 1). Such software may also help also with the triage and patients selection process for endovascular reperfusion therapy by notifying the medical team within minutes [36].

Recent trials have demonstrated the therapeutic potential of thrombectomy even in an extended time window [37, 38]. Patients who can be treated successfully in such an extended time window are referred to as “slow progressors” and it is thought that sufficient collateral blood flow is the key essential to this phenomenon [39]. In the acute stroke set-up, CTA can visualise collateral flow independent of particular arterial territories as compared to DSA. There are numerous qualitative and semi-quantitative scales to score collaterals [40] (collateral scores; CS). A common scoring system was established by Tan et al. [41]. Here, the whole MCA territory is graded from 0 (no collaterals visible) to 3 (no difference to the contralateral hemisphere). However, this score does not take into account anatomic location and functionality and can cause considerable inter-observer variability [40]. As an alternative, anatomic regions like those underlying the ASPECTS can be applied to score collaterals [42] which has been shown to increase the concordance between different readers in comparison the NCCT-ASPECTS [43]. The fact that different scores exist and are used in parallel is mainly due to the fact that there is no “ground truth” against which these scores could be validated. Information from DSA examinations, which is the gold standard for depiction of intracranial vessel, cannot be compared immediately. This is because in CTA contrast agent (CA) is given intravenously and spreads systemically throughout the arterial vasculature and can thereby reach the occluded territory in a retrograde fashion.

As an alternative, validation can be done indirectly by investigating the ability of a particular score to predict the clinical outcome of the patient, appropriate therapy (i.e. vessel recanalization) provided. Consequently, CTA-based scoring of collaterals has proven to be predictive not only for the success of recanalization but also for clinical outcome in several MT studies [44‐46]. Different imaging patterns and patients outcomes dependent on collateral flow are illustrated in Figs. 3 and 4.

It was also demonstrated that the timing of the CTA is crucial for the assessment of collateral flow and that acquisition in the late arterial or early venous phase increases the specificity [47]. This is important since arterial CTA typically aims at the early arterial phase which could lead to an underestimation of collateral flow.

The numerous different scores and the dependence on acquisition techniques as well as individual parameters such as blood pressure or cardiac ejection fraction all can lead to considerable interrater variability in assessing collateral flow on CTA images and standardisation would be beneficial [48]. So far, there are only few reports on automated analysis of CTA images in acute stroke. Regarding automated analysis of collateral flow using CTA images, Boers and colleagues [49] used data from the MR CLEAN trial and compared automatically assessed quantitative CS with those from the trial’s core lab. They found a slightly but non-significantly better outcome prediction by automated CS than by manual scores.

There is a commercial software available from Brainomix called e-CTA® which is embedded in the company’s e-Stroke Suite platform. It determines the Tan score and also gives a percentage of vasculature detected in the affected MCA territory as compared to the contralateral side, similar to the aforementioned approach [49]. It has been evaluated first in the CATS study [29]. Here it was used to score single phase CTA images of 98 acute stroke patients with LVO. The automated CS achieved a high agreement with the consensus score from three experienced neuroradiologists and knowledge of the automated CS also significantly increased the interrater agreement between the three experts. In another study, using single phase CTAs from 235 acute stroke patients undergoing MT, CS from e-CTA were compared with those from two blinded neuroradiologists against an expert-based ground truth. Here, e-CTA again reached a similar level of agreement with ground truth as the professionals [50].

Another application of machine learning algorithms to acute LVO CTA is automated detection of the occlusion. In contrast to CS, where automatization can overcome interrater variability, the advantage is different from LVO detection, since experts are not really challenged. However, less experienced physicians in primary care institutions will depend on the correct detection of an LVO when deciding to refer a patient to a comprehensive stroke centre or not. Technically, such an approach is straight forward as algorithms such as the “region growing” approach can be employed here. There are two reports on this issue so far. In the ALADIN study, Barreira and colleagues used a commercial software employing an AI-based algorithm (Viz LVO, Viz.ai, San Francisco, USA), in 875 patients with 46% LVO and compared the results against an expert-based ground truth [51, 52]. They report an accuracy of 86%, sensitivity of 90.1% & specificity of 82.5 for the software which took under 5 min. Seker et al. [53] used a machine learning algorithm developed by Brainomix in 144 acute stroke patients of whom 73 had an LVO and MT. Here, the algorithm reached accuracy, sensitivity and specificity of 90, 91 and 90%, respectively, when compared against ground truth. The average duration for each analysis was below 1 min. Furthermore, this performance was similar to that of two blinded neuroradiologists.

Automated analysis of CTA using AI-based algorithms can have several advantages in acute stroke imaging: i) interrater variability can be reduced to increase objectivity, ii) less experienced physicians can get support in diagnosing LVO and collateral flow and iii) decision making in stroke treatment can be accelerated. As with human reading, however, timing of the arrival of contrast is pivotal here, too.

In order to derive information from perfusion maps, it is important to define specific imaging parameters. Cerebral blood volume (CBV) is commonly referred to a volume of blood in a given region of brain tissue as per millilitres per 100 g of brain tissue [54]. Cerebral blood flow (CBF) is a certain amount of blood volume passing to a defined volume of brain tissue in a given period of time. This is usually expressed as millilitres of blood per minute per 100 g of brain tissue [54].

Mean transit time (MTT) is the average time (in seconds) a certain volume of blood takes to pass through a given volume of brain [54]. MTT is calculated by dividing CBV by CBF. Time to peak (TTP) is the time it takes an IV-injected bolus of contrast material to reach its peak in a given region of the brain, also commonly measured in seconds.

Physicians need to be aware of major pitfalls that could lead to a falsified or useless perfusion imaging:

In order to evaluate the quality of the perfusion data and to assess the reliability of the post-processing results, both situations need to be taken account for using the arterial input function (AIF) and venous output function (VOF) profiles during imaging post-processing. Computer-assisted imaging algorithms usually provide AIF and VOF profiles automatically. Moreover, “head shaking” head movement can be corrected for by using the midline of the brain. If, however, the patient moves his head within the longitudinal axis (in a figurative sense saying “yes”, nodding), the sections of the brain change the level or imaging slice and voxels usually cannot be reassigned properly during imaging postprocessing.

Over the years, there have been several different approaches to distinguish normal tissue from hypoperfused, viable tissue (‘tissue at risk’) from nonviable tissue (‘ischemic core’) using CT- and MR-perfusion imaging. Correspondingly, numerous thresholds have been advocated [55]. Using CTP, the ischemic core is defined as a region with a substantial reduction in CBF with respect to the healthy contralateral hemisphere (rCBF). A substantial reduction is determined to be present as a reduced rCBF < 30% or 40%. DWI however, is recognised as a gold standard to assess the infarct core based on apparent diffusion coefficient (ADC) thresholds between < 0.6–0.62 × 10–3 mm2/s. (see the Additional file 1). For CTP and MR Perfusion, a delay in time to peak perfusion longer than 6 s (Tmax > 6 s) is considered to be a reliable predictor of the tissue at risk [56‐59]. In a small patient cohort, Lin et al. could demonstrate that CTP and MR Perfusion could be used interchangeably if Tmax = 4 to 6 s measurements were used [60].

The amount of absolute (ml) or relative mismatch between ischemic core and tissue at risk varies within studies [61, 62], but a mismatch ratio of > 1.2 has been used in numerous prospective randomised controlled stroke trials (see the Additional file 1). However, other trials – DEFUSE 3, Solitaire™ FR With the Intention For Thrombectomy as Primary Endovascular Treatment for Acute Ischemic Stroke (SWIFT PRIME) trial – used a tighter mismatch ratio of ≥1.8 [63, 64]. Although the mismatch ratio originates from trials using MRP for patient selection, it is generally accepted to identify patients for reperfusion therapies based on CTP using an rCBF/Tmax mismatch ratio of > 1.2 [64, 65].

Post-processing software estimates perfusion parameters based on above mentioned brain perfusion principles using deconvolution of tissue and arterial signals (e.g. singular value decomposition (SVD) deconvolution method, methods using Bayes theorem) in perfusion imaging (incl. co-registration with DWI images in MRI) with specified thresholds for an automated segmentation and quantification of the infarct core and the tissue at risk [66, 67]. For MRI, in order to reduce false-positive detection of ADC lesions of otherwise healthy tissue automated, infarct core volumetry within the tissue at risk can be used [68].

Because of different post-processing algorithms and thresholds – results differ among commercially available fully automated software even when using identical source data [69, 70]. Physicians need to be aware of this circumstance, especially in difficult decision-making situations or when treating acute stroke patients and perfusion results are somewhere in between against or in favour of a specific treatment. Yet, automated analysis seems to outperform human thresholding and analysis of CTP data [71].

Recently, a different approach has been used to either improve or replace previous software solutions. By using machine learning methods, i.e. deep learning techniques using neural networks, the infarct core volume and tissue at risk can be predicted directly from the CTP or MRP source images. Additional metadata – such as the time parameters and treatment – could further increase prediction accuracy and might even predict infarct growth over time or depending on treatment modality (see the Additional file 1).

There are several vendors, respectively software products for post-processing of CT and MR perfusion images available; some of them are listed in Table 1. Since there are so many software solutions, it is difficult to list all of them and their validation within smaller studies or trials. Vendor-specific software usually has the advantage that thresholds can be chosen at the discretion of the radiologist evaluating the images. This, however, makes it prone to interrater variability leaving quantitative values – if provided by the software – hardly comparable.

The RApid processing of PerfusIon and Diffusion [RAPID] software (iSchemaView) – a commercial software solution – has been used and is currently used in a number of large stroke trials (see Table 2). However, prominent trials using RAPID – EXTEND-IA, DEFUS 3 and DAWN and DEFUSE 3 – only included 70, 182, respectively 206 patients. Furthermore, all these trials did not evaluate the performance of the software. As there are no studies or trials directly comparing their impact on patient outcome depending on the specific imaging post-processing results, it remains unclear whether available software solutions are interchangeable. Such trials, allowing inclusion of patients by multiple vendor products appear to be mandatory to further improve imaging evaluation and make its results more generalizable.