An advantage to the implementation of radiomics in AVMs is that it utilizes routine clinical images obtained in patient management such as CT, CTA, MRI, MRA, and DSA. An imaging protocol is put into place where medical images are obtained from large, established patient databases or ones that are newly created. Ideally, imaging protocols would be described in-depth for reported models. There are currently no universal standardized imaging protocols. Standardizing imaging protocols and/or disclosing current protocols utilized is necessary in order to provide reproducible and comparable results across studies (16). Imaging patients multiple times can also mitigate the variability in radiomics features seen with motion degradation (16, 30).

Once identified and saved within a database, images are stored as their essential components. Segmentation of the ROI/VOI can then be performed manually or with auto-segmentation, which involves the use of automatic or semiautomatic algorithms. A number of studies have been described which utilize machine learning for AVM segmentation in both MR imaging and CT (28, 31, 32). Dice Similarity Coefficient (DSC) is one of the most commonly used performance metrics in the imaging segmentation literature. Unfortunately, this metric is not yet uniformly associated with clinical utility (3, 33). During the process of segmentation, the choice of which voxels, a unit of graphical information that defines a three dimensional space, to incorporate into analysis is determined. This is a crucial step as production of different radiomic features can result from different segmentation techniques (manual vs. semi-automatic vs. automatic) and leads to bias during analysis (16). In order to improve quality of the model, and limit bias, multiple segmentations by different algorithms or physicians can be performed (16).

Qualitative and quantitative information can be derived from medical images. Qualitative features are what is commonly used in radiological vernacular. Today, countless quantitative features can be identified and extracted using high-throughput computing. Feature extraction involves the use of manual image processing or AI to extract a large number of features from medical imaging via mathematical algorithms. After segmentation, individual features can be assessed for correlation to a chosen outcome. Extracted quantitative features can be loosely broken into three categories: shape, signal distribution, and texture (34). Table 4 further categorizes these subgroups. Any combination of these features can lead to the generation of hundreds of variables from a single image (34).

One important variable to consider during feature extraction is the scanner type and vendor. Studies which utilize various scanners must take this into account as it can affect subsequent analyses and may not be externally valid (30). One method to address the uncertainty seen between scanners and vendors is the implementation of phantom studies. These studies use a specifically engineered object which provides an optimal imaging scenario to evaluate and tune performance among different imaging devices (16, 30).

This part of the workflow deals with quality analysis and is often the limiting stage. At this point, radiologic data can be enriched with the addition of other patient data including genetic, biological, and clinical information.

During analysis, all of the information incorporated into the algorithm (radiologic, genetic, biological, clinical) is synthesized and repetitive components are removed. Then, features can be compared and contrasted in order to examine specific outcomes.

Radiomics has been mostly employed in tumor pathology with few studies involving vascular pathology. Current AVM radiomics models have yet to undergo external validation and are not ready for clinical use. Key components of modeling include training, validation, and quality assessment.

Machine learning-based radiomics models must undergo training and validation before clinical use. Performance metrics are most commonly reported as discrimination statistics such as ROC, AUC, accuracy, sensitivity, specificity. There is heterogeneity in how these metrics are reported. Specifically, studies do not uniformly report statistical methods for both training and validation models. Training and validation models should demonstrate consistency in reported statistics (16). There is also confusion in commonly used terminology. Protocols for reporting guidelines in radiomics models and their validation process could lead to improved quality of these studies (14). Lambin et al. (16) described how to perform quality assessment of radiomics models at length. The mean RQS of our studies was 15.92 (range: 10–18). This is comparable to the values of radiomics models in other fields (35–41).

Shi et al. (18) described a hybrid machine learning/deep learning network which utilized temporal and spatial features from DSA videos to diagnose and grade AVMs. This was one of the few studies which studied DSA videos—arguably the most important imaging technique in evaluation of cerebrovascular diseases (42). While some would argue that the diagnosis of AVM is clear without utilizing advanced models, this is not always true in inexperienced centers. Particularly in emergency situations, a tool which aids in diagnosis could lead to faster treatment.

An advantage to using DSA in diagnosis is that it can provide in-depth detail about the cerebrovascular anatomy, including specific arteries feeding the AVM, venous drainage pattern, and perfusion and distribution patterns. Such advantages are highlighted in Zhu et al.’s work,32 which used machine learning analysis of DSA to evaluate hemodynamic differences in ruptured versus unruptured AVMs. This is an advantage over models of AVM diagnosis which utilize 4D-CTA and 3D-RA but do not allow for visualization of vascular perfusion and distribution within the AVM (18, 43–45). Extraction of relevant data from DSA videos had not been proposed prior to the work done by Shi et al. (18) In order to do so, they created a hybrid machine/deep learning model which extracted information from DSA videos by using automated methods to collect information on AVM perfusion and distribution and then combining that with information gleaned from intensity, texture, and wavelet radiomic features taken from static images.

Zhang et al. (19) produced a successful radiomics model for differentiating between AVM-related hematomas versus hematomas from a different etiology using NECT. This is a clinically useful tool as NCET is often the first step in evaluation of suspected intracranial hemorrhage. Distinguishing AVM-related hematomas by the naked eye remains difficult and angiography remains the gold standard. However, angiography can be time consuming and requires the use of contrast agents and patient cooperation. Zhang et al. described their model as a fast, non-invasive method which did not require the use of contrast agents for diagnosis. Their model demonstrated that hematomas resulting from AVM had a larger diameter, coarser texture, and more heterogeneous composition when compared to non-AVM hematomas (19). Their model also outperformed interventional radiologists when comparing measures of specificity and accuracy (19). Distinguishing between AVM-related hematoma and hematomas from other etiologies allows for a faster, more accurate treatment response with targeted utilization of resources.

A significant portion of the work involving ML and DL for AVM has been devoted to methods of delineating the AVM nidus and surrounding structures in preparation for radiosurgery. Specifically for describing methods of auto-segmentation of the ROI/VOI for precise SRS target planning (3, 31, 46). While these studies do not fit the criteria for radiomic modeling, they warrant discussion as they show promise for future studies.

Simon et al. (3) created the first model attempting to generate cerebrovascular anatomical maps in AVM patients. They provided a proof-of-principle study using multiple MRI sequences demonstrating an effective method for producing high-resolution images capable of differentiating not only the AVM nidus, but collateral vessels and surrounding parenchyma for pre-SRS planning. Their model was also able to differentiate those vessels that had been previously embolized. Wang et al. (31) also successfully described this method using CT.

It is currently unclear which AVM patients, particularly those with large or high-grade AVMs, could potentially benefit from DS-SRS (dose-staged SRS) versus VS-SRS (volume-staged SRS). A benefit for DS-SRS is that it allows for multiple sessions of targeted radiation as opposed to a single, large dose (47). Meng et al. (20) extracted radiomic features from stereotactic MR images in order to predict the rate of nidus obliteration (volume reduction velocity) following DS-SRS (dose-stage SRS). They found one radiomic feature (SurfaceVolumeRatio) that was associated with volume reduction velocity. Following DS-SRS, a smaller SurfaceVolumeRatio predicted a higher volume reduction rate. This supported the author’s statement that DS-SRS is more appropriate for AVM niduses which are large and compact as AVM obliteration may be achieved faster (20).

Three of the included studies predicted epilepsy in unruptured AVMs (23, 27, 28). Zhao et al. (27) retrospectively reviewed the database of 2 prospective clinical trials in order to accurately predict epilepsy from TOF-MRAs. Zhang et al. (28) and Lin et al. (23) produced similar results with T2-weighted MR imaging. Early prediction of seizure development in unruptured AVMs could lead to a change in management advocating for more aggressive intervention.

To date, two radiomics models have been reported that assess the risk of rupture in AVM patients, utilizing DSA and CT imaging (21, 24). In young adults, AVMs are one of the most common causes of intracranial hemorrhage (ICH) resulting in a morbidity rate of 30–50% and a mortality rate of 10–15% (18, 48). While the yearly rupture rate of AVMs is about 2–4%, prognostic factors that alter this risk are uncertain.

AVM diffuseness is a prognostic factor of microsurgical resection and has been associated with hemorrhagic risk (29). However, this value can sometimes be hard to quantify and exist along a spectrum. Jiao et al. (29) developed a machine learning algorithm that was capable of determining diffuseness from extracted first-order statistical radiomic features on 3D TOF-MRA. This could help stratify patients before surgical intervention.

Other ML approaches have been described in the prognosis of AVM management. Following SRS, it can take up to 3 years for an AVM to be fully obliterated (46). In this time period, it is important for these patients to remain under clinical surveillance for brain edema and other radiation effects. Meng et al. (25) and Gao et al. (26) described radiomics models utilizing MRI for predicting outcomes of SRS after partial embolization and gamma knife radiosurgery, respectively. Moreover, although not in the realm of radiomics, Yang et al. (32) and Peng et al. (46) both described ML algorithms that were capable of predicting radiation-induced changes following SRS. Using T2-weighted MR images, Yang et al. (32) calculated the volume and proportion of brain parenchyma within the 12 Gy radiosurgical volume (V12) in patients with unruptured AVMs and previous SRS. They found that the volume of brain parenchyma within the 12 Gy radiosurgical volume (V12) following SRS was associated with both early and late radiation-induced changes (32). Similarly, Peng et al. (46) developed an algorithm which was capable of automatically segmenting different regions within the prescription isodense region. This could predict adverse effects in the latency period. Interestingly, radiomics analyses may also target specific outcomes, as Jiao et al. (22) described deep and machine learning methods for predicting motor deficits following AVM resection. These findings underscore the utility of radiomics in postoperative care.

In the era of precision medicine, radiomics has the potential to tailor AVM management as individual patient data can be converted to digital images and mineable data. It is feasible that radiomics will become routine practice in the future as these analyses are intended to be conducted with standard of care medical images (12). Eventually, it is possible that radiomics can personalize AVM management by selecting the most appropriate therapy for each patient and identify more rapidly if it is not working. An area of particular interest would be the use of radiomics to guide treatment of high-grade AVMs, which currently involves considerable debate (49). Radiomics could also assist in the assessment of AVM genetics. The combination of radiological data and genetic data is known as radiogenomics. AVMs represent an interesting pathology for implementation of radiogenomics given recent findings in genetic biomarkers.

Specific to AVMs, the ability of radiomics to handle large volumes of data could lead to both better predictive models describing the natural history as well as treatment response and success. For instance, utilizing quantitative magnetic resonance angiography, flow rates in the feeding arteries and draining veins could be calculated. At this time, it is unknown if these measurements are predictive of rupture or predict a positive or negative treatment outcome (50–52). Due to the large amount of data, it is difficult to determine these relationships manually. Using radiomics, ML and AI could easily be used to study these parameters over a much larger data set.

A large part of the growing field of radiomics is the integration of AI methods into models. It is important to note that the goal of AI is not to replace physicians; rather it acts as a clinical tool. One study showed that a combined model incorporating both a clinical and radiomics model performed better than either model alone (27). In its most basic form, AI is any technology that uses explicit programming codes in an attempt to mimic human behavior. The fact that a code would have to be written for any and every given scenario, makes it rudimentary and non-ideal for AVM management as many thousands of variables could be present given an individual’s anatomy and past medical history. Machine learning (ML) is a subset of AI. It is more sophisticated as it does not rely on explicit programming. Instead, the machine is able to learn the programming rules. However, it is not a fully intelligent method as humans are still defining the features of the programming. One advantage of AI in radiomics studies is the elimination of bias. Evaluation of these films by the operating surgeon or radiation oncologist often introduces bias as that physician already may have a preconceived notion of what they would like to do. Similarly, they may be biased toward one treatment outcome or another. Radiomics would allow the use of objective data to evaluate treatment outcomes and surgical risk, eliminating this type of bias.

Deep learning (DL) is a subset of ML inspired by the brain’s biological neural networks. This technology has the capability to create its own neural networks. When this neural network has multiple layers, it is called a deep neural network. A recent meta-analysis examining DL in neurosurgery showed promising results of DL models (53). However, their review demonstrated that aside from public databases, there was a paucity of readily available data (53). Especially when dealing with large data stores, large-scale data sharing is necessary to promote transparency, create reproducible results, and mitigate bias (53).