Segmentation techniques of brain arteriovenous malformations for 3D visualization: a systematic review

Background Visualization, analysis and characterization of the angioarchitecture of a brain arteriovenous malformation (bAVM) present crucial steps for understanding and management of these complex lesions. Three-dimensional (3D) segmentation and 3D visualization of bAVMs play hereby a significant role. We performed a systematic review regarding currently available 3D segmentation and visualization techniques for bAVMs. Methods PubMed, Embase and Google Scholar were searched to identify studies reporting 3D segmentation techniques applied to bAVM characterization. Category of input scan, segmentation (automatic, semiautomatic, manual), time needed for segmentation and 3D visualization techniques were noted. Results Thirty-three studies were included. Thirteen (39%) used MRI as baseline imaging modality, 9 used DSA (27%), and 7 used CT (21%). Segmentation through automatic algorithms was used in 20 (61%), semiautomatic segmentation in 6 (18%), and manual segmentation in 7 (21%) studies. Median automatic segmentation time was 10 min (IQR 33), semiautomatic 25 min (IQR 73). Manual segmentation time was reported in only one study, with the mean of 5–10 min. Thirty-two (97%) studies used screens to visualize the 3D segmentations outcomes and 1 (3%) study utilized a heads-up display (HUD). Integration with mixed reality was used in 4 studies (12%). Conclusions A golden standard for 3D visualization of bAVMs does not exist. This review describes a tendency over time to base segmentation on algorithms trained with machine learning. Unsupervised fuzzy-based algorithms thereby stand out as potential preferred strategy. Continued efforts will be necessary to improve algorithms, integrate complete hemodynamic assessment and find innovative tools for tridimensional visualization.


Introduction
Brain AVMs (bAVMs) are complex vascular lesions. Tailoring management for each patient is challenging and demands an accurate knowledge and understanding of the angioarchitecture of the malformation. bAVMs represent a relevant cause of secondary intracerebral hemorrhage, with a risk of rupture of approximately 1% yearly that on average increases fivefold after rupture [1][2][3].
Brain bAVMs are composed of feeding arteries and draining veins entangled in a nidus, without an intervening capillary bed [4]. Visualization of these structures is vital in the understanding of the angioarchitecture. Comprehension of flow direction and amount is necessary to grasp the hemodynamic effects. Determination of vessel positions in relationship to the nidus and brain structures is fundamental to optimize treatment strategies. The best imaging tools for the characterization of bAVMs are digital subtraction angiography (DSA) and magnetic resonance angiography (MRA) [5]. Both imaging modalities play a role in the diagnostic and perioperative management of these lesions, highlighting bAVMs' features that drive their therapeutic management [6][7][8]. In neurosurgery, based on clinical and imaging findings, the severity of bAVMs, their predicted risk of rupture and necessity for intervention are mostly rated according to the classical Spetzler-Martin classification and its supplementation provided by Lawton et al. in 2010, demonstrating a stronger correlation with surgical outcomes [9][10][11].
Visualization of feeding arteries, nidus and venous drainage system is vital in the understanding of bAVM angioarchitecture and related hemodynamics: to tailor a therapeutic strategy, it is important to determine vessel positions and flow directions and to distinguish them depending on their nature and role. An accurate understanding of the angioarchitecture of bAVMs can be achieved by segmentation [3,[12][13][14][15][16]. Segmentation is intended as manual delineation of structures on a Digital Imaging and Communications in Medicine (DICOM) series to create its 3D shape. With segmentation, size and shape of cerebral structures can be measured to determine their spatial characterization and to plan a precise therapeutic intervention [16,17]. Modern visualization techniques such as augmented reality (AR), 3D screens, as well as mixed reality (MR) and virtual reality (VR) have also been implemented to show segmentation results [18][19][20][21][22].
Several publications have reviewed vessel extraction techniques [15,[23][24][25]. They have helped to classify segmentation and visualization strategies and show the diversity and limitations of the used methods. Nonetheless, a major lack in the current literature is the presence of a gold standard technique shared by multiple centers worldwide to address segmentation of bAVMs, as well as the integration of hemodynamic information to characterize not only the morphology, but also the physiology of these complex lesions. The aim of this systematic review is to provide an up-to-date collection of on 3D methods to study the angioarchitecture of bAVMs for pretreatment planning and during embolization, surgery or radiosurgical treatment, as well as their integration to hemodynamic information and augmented reality rendering.

Materials and methods
A systematic review was performed using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [26]. Two reviewers (EC and TF) screened records independently, and disagreements at any stage were resolved by discussion and consensus. Two additional records were identified through reference search. Studies were excluded when considered beyond the scope for the aims of the present analysis, and/or when their outcomes were not of interest. The critical appraisal of the included studies was performed by means of risk of bias score as shown in Table 1.

Search strategy
The PubMed, Embase and Google Scholar databases were searched to identify eligible papers. The query was performed using the Boolean operators "AND" or "OR" and database-related filters to maximize the chance to identify articles focusing on segmentation strategies and their 3D visualization specific of bAVMs. The string (('cerebral arteriovenous malformations') AND ('brain arteriovenous malformations') AND segmentation OR 'tridimensional visualization' OR '3D' OR 'skeletonization methods' OR 'augmented reality') was entered. The most recent search was performed on February 18, 2022.

Selection criteria
Articles were included if all the following criteria were met. 1) Studies published after 1997; 2) studies analyzing specifically the angioarchitecture of brain arteriovenous malformations; 3) a 3D segmentation technique on top of source data as a mean to study angioarchitecture; 4) English, Italian, French or German language; 5) studies integrating the segmentation outcome with augmented reality technology.

Statistical analysis
The descriptive statistical analyses were performed using IBM SPSS Statistics 25. Data were presented as numbers and percentages, and medians with IQR. Heterogeneity was tested by Chi-square test (significance level: p-value < 0.01).

Results
A PRISMA flowchart is displayed in Fig. 1. A total of 7212 publications were screened, 37 full-text articles were assessed for eligibility and 32 studies were included in this review.
The most utilized imaging input in the automatic segmentation subgroup was MRI (10/20, 50%).

Segmentation variability
A Chi-square test of homogeneity was performed to examine the methodological variability of the collected studies. The homogeneity between the segmentation strategies and the reported outcomes, namely segmentation of all bAVM components, segmentation of the nidus only or other outcomes (e.g., volume), was not significant, with a Chi-squared = 0.13 (p-value 0.99).

3D visualization
Four studies (12%) described an integration of their segmentation strategies with mixed reality for intraoperative guidance under the microscope [37,47,55,56]. While integration with AR was positively endorsed, the future necessity to integrate hemodynamic considerations to the AR rendering to provide more useful and precise intraoperative information was also suggested [18]. None of the included studies aimed to a rendering of the segmentation outcomes with VR.

Clinical value and blood flow analysis
In 19 studies (58%), the final goal of segmentation was an accurate and more precise planning and intraoperative management of bAVM [25, 27-33, 35, 38, 43, 45, 48, 50-52, 54-56]. Eleven studies (33%) aimed to a better delineation of bAVM components to achieve a more effective and safer radiosurgical treatment [20, 21, 34, 36, 37, 39-42, 49, 53]. Three (9%) of the included studies set embolization as the final goal of segmentation and better angioarchitecture visualization [44,46,47].Only one research group aimed and achieved the inclusion of hemodynamic considerations to 3D bAVM visualization. They could provide temporal information on blood flow of the vessels in the close proximity of the nidus by calculating a time-to-peak parameter map and registering it to TOF-MRA sequences [32,33].

Discussion
The present study represents a systematic analysis on segmentation techniques of bAVMs for 3D visualization and accurate angioarchitectural study. This systematic review focusses on 3D bAVM segmentation and visualization strategies, which have been published from 1997 to 2022 on the characterization of the angioarchitecture of bAVMs.
Although most studies use MRI as input and a tendency was shown in the direction of automatic machine learning trained algorithms, this review shows how significantly variable the possible methods of segmenting a bAVM are, and with this the absence of a gold standard. Due to methodological variety of manual, semiautomatic and automatic segmentation techniques and their outcomes, the results should be interpreted with caution. Intrinsic biases of included publications cannot be ruled out. The average segmentation duration was evaluated using very limited data from only a few studies. Because of the great variability of the data collected in the present systematic review, a descriptive statistical analysis was performed. Given the technical and preclinical nature of most of the included studies, very few of them documented sensitivity and specificity of the segmentation method. Therefore, the scarcity of statistical data made the performance of a pooled analysis impossible.
This review describes a tendency over time to base bAVM segmentation on algorithms trained with machine learning, especially deep learning, because manual segmentations are prone to human error and interindividual assessment and are labor intensive [28]. Algorithms trained with machine learning stem from two principles: supervised and unsupervised learning. Supervised learning algorithms, such as support vector machine, have the disadvantage of relying on manual segmentations. In addition, the variability of scans should be large, because algorithms trained on scans from 1 center are rarely easily extrapolated. Unsupervised learning does not require initial information, and algorithms based on this technology are generally fast. In the present analysis, fully automated segmentation via unsupervised learning has good results using fuzzy-based methods. This approach is based on the computation of vesselness filter and maximum parameter images on MRI-TOF sequences, providing a highly precise delineation of large as well as fine vessels [32,36].
The deep-learning algorithm developed by Forkert et al. describes the extra advantage of producing surface models of the vascular system, which can be not only visualized but also manipulated in 3D [28,32]. In the settings of 3D segmentation, operability of surface models differentiates them from outcomes of computational heavier volumetric models, that in general are more detailed and look sharper than surface models but do not allow any interaction and are mostly too heavy to visualize in a head-up display without streaming [57]. In view of extensive integration of medical imaging with mixed reality or virtual reality in the future, the use of surface models versus volume models as a possible output of algorithms in highly complex 3D lesions as bAVMs should be further investigated.
In a medical world moving toward personalized medicine, segmentation strategies and 3D imaging visualization techniques are increasingly gaining popularity. Pre-and intraoperative delineation of complex anatomical entities like bAVMs with these technologies provides important clinical advances [58]. First is a more precise understanding of bAVM angioarchitecture and anatomical relationships with the surrounding structures [37,38]. Therefore, a more accurate and individualized therapy planning and the possibility to achieve more efficient patient management and potentially better clinical outcomes [56]. Only one research group incorporated flow in 3D bAVM segmentation. However, flow is extremely important in assessing brain AVM (re-)rupture risk analysis [59]. Transcranial Doppler (TD), DSA and quantitative MR-Angiography (QMRA) [60,61] have been described as useful techniques to analyze flow in bAVM [22,[62][63][64]. The integration of these techniques in 3D bAVM segmentation should be a future goal.

Conclusion
A golden standard for 3D visualization of bAVMs does not exist. This review describes a tendency over time to base segmentation on algorithms trained with machine learning. Unsupervised fuzzy-based algorithms thereby stand out as potential preferred strategy. Continued efforts will be necessary to further improve algorithms, integrate complete hemodynamic assessment and find new innovative tools for tridimensional visualization.
Funding Open access funding provided by University of Zurich. This work is part of "SURGENT" under the auspices of University Medicine Zurich/Hochschulmedizin Zürich.

Declarations
Conflict of interest Tristan van Doormaal is Co-founder and CMO of Augmedit bv, an augmented reality company.
Ethical approval Ethical approval will not be required because this study will retrieve and synthesize data from already published studies.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.