Machine Learning Approaches for Segmentation of Cardiovascular Neurocristopathy Related Images

Cardiovascular neurocristopathy is associated with abnormal migration and development of neural crest cells, impacting the neural and the human cardiovascular system and leading to diseases such as cardiomyopathy, aortic disease, and aortic valve dysfunction. With advancements in biomedical imaging tools, efforts are made to understand the underlying causes of cardiovascular neurocristopathy and develop new diagnostic methods, especially using machine learning or specifically its sub-branch deep learning models. This article provides a systematic survey of the literature related to machine/deep learning-based segmentation of the diseases mentioned above in computer tomography (CT), magnetic resonance imaging (MRI), X-rays, and echocardiogram (Echos) images. The review identified gaps and provides future directions, such as the need for better interpretable and explainable AI models, addressing the lack of publicly available datasets, standardizing the result reporting procedure for better repeatability of the result, and the development of standard performance measurement metrics. The general conclusion suggests that there is a need for multimodalities, multimodel, high-quality data sets, and open-source disease-specific dataspaces that will help develop trustworthy deep learning models that could be implemented in imaging devices/tools and provide medical-grade segmented outputs that will augment and speed up clinician decision making.


I. INTRODUCTION
In recent times, clinical and therapeutic needs have been increasing.According to the World Health Organization (WHO) report published in 2020, cardiovascular disease (CVD) is the leading cause of death in the world, responsible for 16% (8.9 million) of worldwide deaths. 1 As cardiovascular disease continues to be the leading cause of death worldwide, 2 the need for a timely and accurate diagnosis has never been more important [1].Neurocristopathy is a pathological The associate editor coordinating the review of this manuscript and approving it for publication was Derek Abbott . 1 https://www.who.int/news-room/fact-sheets/detail/the-top-10-causesof-death 2 https://www.cdc.gov/globalhealth/healthprotection/ncd/cardiovasculardiseases.htmldisorder that occurs due to an abnormal development and/or migration of neural crest cells, during embryonic development [2].During embryonic development, the migration, abnormal specification, differentiation, or death of neural crest cells results in different neurocristopathies.Various diseases such as thyroid, skin pigments, heart and craniofacial abnormalities, tumors, and digestive tract dysfunction are some examples of neurocristopathies [3].In this paper, we will focus on some cardiovascular neurocristopathy , such as cardiomyopathies, stroke, bicuspid aortic valve diseases and aortic vessel diseases.
Cardiovascular neurocristopathy is a genetic disorder that affects the development of the human cardiovascular system [4].The formation of many different organs and tissues, including the heart and blood vessels, is dependent on the controlled migration and differentiation of neural crest cells [5].Cardiovascular neurocristopathy occurs when gene mutations cause abnormal deaths of neural crest cells and reduce the performance of tissues/organs (for example, the heart) [6].
The aorta is the largest blood vessel and is responsible for oxygen circulation within the body [7]. Figure 1 shows the anatomy of the aorta and its branches. 3From the figure we can see that the aorta gives rise to several branches, including: 1) Ascending Aorta: This portion of the aorta carries blood upward and then curves to form the aortic arch.2) Aortic Arch: The aortic arch is a curved segment of the aorta that gives rise to several major arteries that supply the head, neck, and upper extremities.These branches include the brachiocephalic artery (which further divides into the right subclavian and right common carotid arteries), the left common carotid artery, and the left subclavian artery.3) Descending Aorta: The aorta continues downward as the descending aorta (region below the left subclavian) and is divided into the thoracic aorta (supplying the chest area) and the abdominal aorta (supplying the abdominal region and lower extremities).Efforts are being made to better understand the underlying causes of cardiovascular neurocristopathy and develop new approaches to the diagnosis, treatment, and prevention of such disorders, especially focusing on machine learning and deep learning models [8], [9], [10].Untreated cardiovascular neurocristopathy conditions have severe effects on human health and can cause cardiomyopathy, stroke, aorta, and aortic valve dysfunction diseases [11].Each of the mentioned conditions is defined as follows.
Cardiomyopathy is a disease that affects the shape, thickness, and/or size of the heart muscle [12].Cardiomyopathy can be divided into four subtypes [13], which are itemized below: 1) Dilated cardiomyopathy: where the heart muscle wall becomes thinner and stretched.2) Hypertrophic cardiomyopathy: where the heart muscle wall thickens.3) Restrictive cardiomyopathy: A set of changes occurs in the heart that limits its ability to expand or contract to pump blood into/out of the heart.4) Arrhythmogenic right Ventricular Cardiomyopathy (ARVC): In this condition, fatty fibrous tissue replaces normal heart tissue, causing abnormal heart rhythms.Stroke may occur due to a blood clot reaching the brain as a result of a tear in the wall of the large blood vessel (in the neck), which is known as a dissection of the cervical artery [14].Bicuspid aortic valve diseases could be classified into two classes: 1) Bicuspid aortic disease with aortopathy: where an aortic valve contains only two cusps (or flaps) instead of three with a weakened aorta.2) Bicuspid aortic disease without aortopathy: an aortic valve that contains only two cusps (or flaps) instead of three without a weakened aorta.Whereas, symptoms of the aortic vessel disease include: 1) Aortic dissection: where there is a tear in the inner layer of the aorta.2) Aortic aneurysms: where a balloon-shaped bulge is formed in the aorta.3) Aortic hematoma: where the pool of clotted blood forms an organ, tissue, or body space.Figure 2 illustrates the anatomy of different cardiovascular neurocristopathy diseases. 4he current literature presents the use of different imaging and sensory techniques as well as health records for the diagnosis of these diseases [4], [15], [16], [17].In imaging, magnetic resonance imaging (MRI) and computerized tomography (CT) are the most widely used modalities.The echocardiogram (Echo) is used for the assessment of associated cardiac abnormalities while X-rays are the most widely used for screening.The electrocardiogram (ECG) and the photoplythesmogram (PPG) are included in the sensory methods.Some studies use electronic health records (EHR) to collect data or identify those already diagnosed with the disease.Diagnosis of diseases such as CV neurocristopathies require often history taking, imaging, and bio-markers and it requires specific criteria to establish the diagnosis.These are called diagnostic criteria.
This paper focuses on reviewing the available literature on the segmentation of aforementioned diseases where imaging data is used as input for analysis.

A. DIFFERENTIATING/COMPARING MODALITIES FROM A COMPUTER SCIENCE PERSPECTIVE
From the perspective of computer science, each of the imagining modalities is different in terms of sensors providing different information about the heart.
1) X-ray: An X-ray is a black-and-white image that captures the outline of the heart and the major blood vessels.X-rays are useful for the detection of large structural abnormalities but do not provide details on heart function [18].X-rays are the building block of the computer tomography image and are usually not used alone for the detection and segmentation of cardiovascular diseases [19].2) Computer Tomography (CT): Computer tomography is a 3D image that captures details of the size, shape, and function of the heart, as well as any narrowing or blockage of the arteries [20].CT images provide accurate information for the diagnosis, segmentation, and monitoring of cardiovascular diseases.However, it uses ionizing radiation (X-rays), which is harmful in high doses and therefore not suitable for certain vulnerable patients [21].3) Magnetic Resonance Imaging (MRI): Magnetic resonance imaging is a high-resolution image captured using a magnetic field, rather than radiation, and a sophisticated machine to capture information about the heart and arteries [22].Magnetic resonance imaging provides detailed images of the structure and functionality of the heart, including areas of narrowing, blockage, and leakage in the heart, as well as in the arteries.However, MRI scans are time-consuming, expensive and not suitable for patients who are claustrophobic or who have implants [23].4) Echocardiography (Echo): An echocardiogram captures sound waves that bounce off the heart and its blood vessels to provide information about its structure and function, including how the heart pumps blood and the extent of any narrowing or blockages of the arteries [24].Echoes are inexpensive, noninvasive, and can be performed quickly and at the bedside.
The images produced through this method are less detailed and low-resolution compared to CT and MRI, therefore, they are only used if continuous monitoring of vulnerable patients is required [25].The image quality and information within the images obtained through each modality are different and have a significant impact on the development of the segmentation model.The image quality directly affects the training and validation accuracy of segmentation algorithms in identifying and delineating anatomical structures.Higher image quality enhances precision while lower-quality image introduces challenges and errors.Furthermore, the information contained within the images also varies across modalities.Some modalities (such as CT and MRI) provide more detailed anatomical information while some modalities (X-ray and Echos) provide only functional/physiological information.
Understanding the different imaging modalities from the point of view of a computer scientist is essential for the development of better diagnostic tools and tools that could be utilized to efficiently analyze and interpret medical images.Computer scientists must consider the above-mentioned factors to select appropriate imaging modalities and develop robust segmentation models that ensure optimal performance and reliable results.

B. CHALLENGES
The traditional imaging techniques/tools have significant challenges, which limit the clinical and prognostic ability of these tools [26], [27], [28].Some of these challenges are as follows: 1) Inter-and Intra-observer variability: Interobserver variability refers to variability in the results when different doctors/physicians interpret the same image, whereas intraobserver variability refers to variability in the results when the same doctor/physician interprets the same image multiple times.In other words, the ground-truth information available for training will not be the same or accurate.2) Delayed reporting on image results: Time-consuming exams can delay the diagnosis and treatment of patients.In some cases, they may have to wait hours and/or days to receive results, which leads to anxiety and uncertainty.3) Suboptimal image quality: The low-quality images limit the clinical validation of cardiovascular images.
It is difficult to see certain areas of the heart in lowquality images, leading to misdiagnoses or inaccurate assessments of the severity of the disease.The suboptimal image quality also hinders the development of a segmentation model due to limited visual information for training.4) Operator fatigue: Performing multiple exams and long shifts in a day can be physically and mentally tiring, resulting in possible errors or/and inaccurate results.5) Radiation exposure and contrast agents: Repeated exposure to ionizing radiation over time and the use of contrast agents to enhance the resolution of a certain structure could increase the chances of developing cancer in certain patients.Furthermore, the contrast levels of the images are also varied for different devices/machines.Thus, a model trained on a certain level of contrast images might not work on images taken through another machine, due to the contrast difference.6) Inter-vendor variability: Inter-vendor variability in certain imaging parameters such as myocardial strain imaging and perfusion images might limit the generalizability of ML algorithms across all vendors and scanners.With continuous technological advancement in the field of cardiovascular imaging, researchers are motivated to develop different analysis tools to overcome the challenges/limitations mentioned above [29], [30], [31], [32], [33], [34], [35].This will help doctors/clinicians diagnose the disease accurately and design better treatment plans for their patients [36], [37], [38].The application of artificial intelligence (AI), especially machine learning (ML) and deep learning (DL), in the field of medical imaging has been growing rapidly in recent years [39], [40].Traditional ML algorithms are fed manually hand-crafted features, while DL algorithms automatically learn the same features for analysis [36].
This work is designed to review the available literature on the segmentation of the aforementioned diseases using imaging data with the help of ML and DL methods.The review covers different aspects of diagnosis using different imaging techniques and preventive care that no other single review covers.Table 1 provides a comparison of the proposed review with other recently published review articles.
The rest of the review article is organized as follows: Section II describes the search terms, the inclusion criteria, and the PRISMA flow chart.Section III provides a detailed review of the available literature on the different methodologies used for each disease and the advantages/limitations of each modality.It also discusses the clinical interpretation of the data and different ML / DL methods used for the segmentation of the aforementioned diseases.Section IV illustrates the findings and shortcomings learned from the literature review, while conclusions and future directions are provided in Section V.

II. SEARCH METHODOLOGY
This review is in accordance with the PRISMA guidelines [45], a preferred reporting process for systematic reviews and meta-analyses, illustrated in Figure 4.The literature searches were performed on the following platform: • PubMed, Google Scholar, IEEE Xplore Digital Library, Elsevier, ScienceDirect, and Wiley online library.
The search terms were the combination of two primary terms (machine learning, deep learning, segmentation) and a maximum of one secondary term (Dilated cardiomyopathy, Hypertrophic cardiomyopathy, Restrictive cardiomyopathy, Arrhythmogenic right Ventricular Cardiomyopathy, aortic dissection, aortic aneurysms, aortic hematoma, bicuspid aortic valve disease with aortopathy, bicuspid aortic valve disease without aortopathy and cervical artery dissection).
Initially, 750 studies were identified, including both journal and conference publications.After reading the title and abstract, 100 articles were selected.Studies excluded after full-text evaluation included nonhuman studies, that had insufficient data or modality used other than MRI, CT, X-rays and/or Echo.A total of 48 articles were shortlisted for review after a full-text assessment.

III. DEEP LEARNING TOOLS FOR CARDIOVASCULAR NEUROCRISTOPATHY SEGMENTATION
Deep learning is a sub-branch of machine learning that utilizes structured and unstructured data to learn abstract representations [46].These algorithms/models are advanced versions of neural networks with specialized layers and special structures for processing the neurons, allowing them to handle any data source and extract useful information [47].Regarding cardiovascular neurocristopathy image processing, the convolutional layers with different numbers of kernels, play a critical role in the ability of algorithms to generate various feature maps that can help in identifying specific/hidden patterns such as edges, colour gradients, and shapes within the input image/video [48].Subsequently, these learned patterns are used to perform segmentation analysis, that is, to detect specific objects or structures through pixel-level labelling [49].Some common algorithms used for image segmentation include Convolutional Neural Networks (CNN) and their variants such as U-net, auto-encoders (AE), and generative adversarial networks (GAN) [50], [51].
The paper although mention both ML and DL, but for segmentation purpose, mainly the DL models are the focus because of their automatic nature of learning discriminative features and recent rapid increase in performance.Table 2 summarizes the performance metrics used in the reviewed studies along with the rationale and characteristics of each metric while the most frequently used DL and ML algorithms for medical imaging segmentation are tabulated in Table 3.
In terms of performance metrics (Table 2), the majority of reviewed studies used the dice score and the Hausdorff distance as primary performance measuring metrics along with a few other supporting metrics, such as sensitivity,  specificity, Intersection-over-Union (IoU), correlation score, and average symmetrical surface distance to determine the efficacy of their segmentation model.These metrics are commonly used in the medical field as they provide valuable insights into different aspects of the segmentation model's performance.The Dice score and IoU evaluate the spatial overlap between ground truth and predicted segmentation, which is important for the delineation of the anatomical structure [52].The Hausdorff Distance estimates the maximum deviation between ground truth and predicted segmentation and is useful in identifying segmentation outliers [53].The Sensitivity and Specificity metrics help in assessing the model's ability to correctly identify positive and negative cases [54].The Correlation Score measures the relationship between the ground truth and predicted segmentation, which can be informative for quantitative measurements [55].Finally, the Average Symmetrical Surface Distance quantifies the average distance between ground truth and predicted segmentation providing boundary accuracy and smoothness of segmentation [56].
The table 3 summarizes the modality, performance metrics, sample size, data availability, and DL / ML algorithms implemented in 24 reviewed articles.One of the most significant challenges in segmenting and treating cardiovascular neurocristopathy disorders is the limited availability of open source publicly accessible data set on the disease, as evidenced by Table 3.Only one out of 24 studies (Guo et al. [57]) has provided the data set on which they worked.Furthermore, nine other studies ( [58], [59], [60], [61], [62], [63], [64], [65], [66]) have mentioned that the data set could be provided to researchers on a reasonable request.All other remaining 18 studies (66.7%) have not noted the availability of their dataset.
As mentioned in I-A, CT and MRI are 3D imaging techniques that provide high-resolution, detailed, and sophisticated images, making them the more trustworthy modality for segmenting cardiovascular diseases.Some studies such as Asif et al. [63] and Qin et al. [67] used CT images for segmentation and compared their results with MRI data.Asif et al. performed segmentation on CT images and assessed segmentation accuracy using MRI as a reference standard collected from 84 patients (45 men and 32 women who underwent CT and MRI scans).The authors reported an AUC of 0.714 and a segmentation accuracy of 68.9% for males and 71.9% for female patient images.Qin et al. performed a semi-automated segmentation on CT and MRI data collected from 161 patients (2576 segments) with cardiomyopathy.They calculated 1514 myocardial fibrosis (MF) features and applied radiomics to the extracted features.The reported result showed the usefulness of radiomics in the segmentation and detection of MF using CT images (and compared to MRI).The authors reported that the AUC for the training and testing cohorts was 0.81 and 0.78 for the segment-based analysis, respectively.
Learning models for disease segmentation using medical images like CT scans, MRI, X-rays, and echocardiograms present several challenges due to the complexity and importance of the task.The following subsections summarise some of the challenges and potential solutions as well as provide review of different deep/ML algorithms used for cardiovascular neurocristopathy diseases with individual image modalities (CT, MRI, X-Rays and Echos) along with the performance metrics used for their evaluation, the sample size used, and open-access availability of the dataset.

A. CHALLENGES AND POTENTIAL SOLUTIONS IN IMAGE SEGMENTATION
This subsection outlines certain difficulties encountered when working with medical images and offers insights into potential solutions aimed at creating an AI-powered clinicalgrade medical image segmentation device for early detection and prognosis applications.Limited Data: Annotated medical image datasets can be limited in size due to privacy concerns and the difficulty of obtaining expert annotations.To mitigate this challenge, data augmentation techniques can be employed to generate additional training samples from the existing data.Transfer learning, where models pre-trained on large datasets are fine-tuned on medical data, can also help when data is scarce.Annotation Quality: Annotations for medical images need to be highly accurate, and inconsistencies or errors in annotations can lead to model biases.To address this, expert radiologists should review and verify annotations to ensure accuracy.Interpatient Variability: Patients can have significant anatomical and pathological variations, making it challenging for models to generalize.To enhance model generalization, data augmentation techniques that simulate variations can be used.Domain adaptation methods can also be employed to adapt models to specific patient populations.Interpretability: Medical professionals often require models to provide interpretable results and explanations for their predictions.To meet this Collaboration with medical professionals and institutions to conduct clinical trials and validation studies is essential to demonstrate the model's effectiveness and safety in realworld scenarios.Real-time Processing: In a clinical settings, real-time or near-real-time processing of medical images is crucial.To achieve real-time performance, models should be optimized for inference speed, and hardware acceleration options (e.g., GPUs, FPGAs) can be explored.
Addressing these challenges necessitates a multidisciplinary approach involving machine learning experts, medical professionals, data scientists, and regulatory experts to ensure the safe and effective deployment of disease segmentation models in clinical practice.

B. LEARNING MODELS FOR DISEASE SEGMENTATION USING COMPUTER TOMOGRAPHY (CT) IMAGES
In the literature, most studies use Unet models to segment cardiovascular neurocristopathy diseases using CT.Of 15 studies, 8 studies implemented UNet models for segmentation while 2 studies ( [82] and [83]) reported the use of GANs for segmentation tasks.The review of the literature indicated that the use of neural networks (especially convolutional neural networks) is also a popular method, frequently implemented for disease-related segmentation tasks.
Regarding performance comparison, no direct comparison of the model's segmentation capability could be achieved because different studies calculated different performance metrics.The most reliable metrics, as calculated in most studies, are sensitivity, specificity, dice score, and Hausdorff distance, along with some other metrics such as accuracy, the Jaccard index, and the calculation of different error values.Figure 5 demonstrates an example of a segmentation pipeline implemented to segment the true and false lumen in a CT scan image.(0) is the original source image, (1) shows the initial step of the segmentation algorithm that accepts the axial image and performs the localization of the aorta, (2) the centreline identification algorithm is used to derive the aortic centerline, (3) then multiplanar reformations are generated orthogonal to the centerline, (4) and true lumen, false lumen and background is segmented using a segmentation algorithm, (5) the segmented lumen is finally superimposed back on to the axial plane image.
Table 4 summarizes the reviewed literature that performed CT image segmentation using different DL tools.Only two studies, [61] and [85], reported the availability/sharing of the CT image datasets on request.

C. LEARNING MODELS FOR DISEASE SEGMENTATION USING CARDIAC MAGNETIC RESONANCE IMAGING (CMRI)
Most studies in the literature use Unet models to segment cardiovascular neurocristopathy diseases using CMRI/MRI.Figure 6 illustrates an example of automatic left ventricle (LV) segmentation in MRI scans using AI.The subfigures a,b, and c are the MRI scans of patients without cardiomyopathy, with DCM and with HCM, respectively.The outer edge of the LV is circled in green colour, the red colour circle is an intimal contour of the LV while the yellow colour circle is the intimal contour of the right ventricle.
Table 5 summarizes the reviewed literature on learning models for the segmentation of cardiovascular neurocristopathy disease using DL models, UNet, CNN, and DeepV3+ on the MRI dataset.A dataset of MRI scans of three studies (that is, [59], [60], [94]) could be provided upon reasonable request to the authors.

D. LEARNING MODELS FOR DISEASE SEGMENTATION USING X-RAYS AND ECHOCARDIOGRAPHY (ECHO) IMAGES
Limited literature was found on the use of X-rays and echocardiography images for the segmentation of cardiovascular neurocristopathy diseases.One reason could be that frequent exposure to X-rays is harmful to human health.Moreover, echos and X-ray images are of low quality and could only be used for structural segmentation, resulting in relatively less detailed analysis.118310 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
where R is the cardiothoracic ratio, D L is heart distance from the left while D R is heart distance from the right of the central vertical line.M is the maximum horizontal distance between the left and right boundary sides of the lungs, as shown in Figure 7. Figure 8 shows an example of transesophageal echocardiography (TTE) segmentation.The figure shows the segmentation of the aortic valve, right atrium/right ventricle, left atrium, right ventricular outflow tract and main pulmonary artery.Subfigure (a) is the original TTE image used as an input in the AI model, (b) is the annotated ground truth image, (c) is the model's predicted labels and (d) is the predicted annotations overlayed on the original TTE image.
Table 6 summarizes the literature found on x-rays and echo images for the segmentation of diseases using DL models.The literature reports results using variants of neural networks as DL models.The performance metrics used are accuracy, precision, recall, and F1-score, while some used dice score and IoU.Only one out of 6 studies (that is, [77]) reported the availability of the analyzed dataset on request.

E. IMPLEMENTATION OF DEEP LEARNING MODELS
To build an end-to-end DL model, some open-source informatics resources have been developed, including Keras, TensorFlow, and PyTorch [106].Keras is a Python/R library that provides high-level functions to build a stack of consecutive layers to form Theano models.Keras library allows the user-friendly network definition, optimization, and effective evaluation of multi-dimensional mathematical expressions.TensorFlow is also a Python/R library developed by the Google team.TensorFlow uses data flow graphs for efficient data processing.Lastly, the PyTorch library, developed by the Meta AI team, works with Python.It is designed to improve the model building and overall data processing speed [43].Table 7 presents a summarized comparison of the above-mentioned libraries in terms of their architecture, script written, dataset sizes, API levels, do they have trained models, debugging capabilities, and speed performance.
From the reviewed literature, it can be concluded that CNN and Unet (and their variants) are the most frequently used deep learning algorithms.Each algorithm, along with its advantages and limitations, is described as follows: Other elements in CNNs include batch normalization and dropout layers to enhance stability.Architectural modifications like dilated convolutions capture multi-scale features, crucial for certain segmentation tasks.Transfer learning with pre-trained CNN models like VGG or ResNet, fine-tuned for segmentation, has proven effective, especially with limited data.
Advantages of CNNs for segmentation include their adaptability, feature learning capabilities, and fast inference times.However, they may struggle to capture fine details and require substantial labelled data.Overfitting is also a concern, necessitating the use of regularization techniques.Finally, the Interpretability of CNN results can be challenging due to their complex architectures.

2) UNET AND ITS VARIATIONS
U-Net is a specialized architecture specifically designed for image segmentation tasks, with a unique encoder-decoder structure featuring skip The encoder captures high-level features from input images, while the decoder recovers spatial information.The skip connections, a distinctive feature of U-Net, play a crucial role in preserving spatial information during the upsampling process, contributing to accurate segmentation.Common loss functions used for U-Net and its variants include the Dice coefficient or crossentropy, while optimization often involves techniques like Adam.
U-Net excels in biomedical and general image segmentation tasks, with skip connections enhancing its ability to capture fine details.Variations of U-Net, such as Attention U-Net and ResU-Net, have been developed to further improve performance through attention mechanisms and residual connections.Nevertheless, U-Net and its variations are specialized for segmentation and may not perform optimally in other computer vision tasks.Creating accurate labelled segmentation masks can be labour-intensive and require domain expertise.Overfitting remains a concern, particularly in scenarios with limited data, and complex or noisy images can challenge U-Net's segmentation performance.

IV. FINDINGS
Table 3 summarizes the learning models implemented in the field of cardiovascular neurocristopathy, while Table 4, Table 5 and Table 6 summarize the learning models implemented in the field of cardiovascular neurocristopathy segmentation using CT, MRI, X-rays and echocardiographic images, respectively.The findings of this review article are as follows.

A. STANDARD PERFORMANCE METRICS AND RESULT INTERPRETABILITY
It can be observed from the tables that some studies implemented the same algorithms but reported the segmentation performance using different performance metrics.This makes a fair comparison of these algorithms more complicated, especially if analysed for shortlisting to the most specific and sensitive algorithm that could be implemented in the field of medicine for accurate segmentation purposes.Therefore, it is necessary to standardize performance metrics for ease of comparison and to determine the most accurate segmentation algorithm for each, as well as the most useful imaging modalities.

B. AVAILABILITY OF OPEN-ACCESS DATASETS, ANNOTATION, AND INTERPATIENT VARIABILITY
The review also revealed that there is limited access to publicly available open-access datasets.Only 1 out of 32 studies [57] reported that their data set was open access, while 6 out of 32 studies [59], [60], [61], [77], [85], [94] reported that the data is available to researchers on reasonable request.The limited availability of open-access medical data is subject to privacy and data security.One of the solutions could be the development of standards for data anonymization and following GDPR rules for data sharing across continents.

C. REPRODUCIBILITY OF RESULTS AND HIGH-QUALITY DATASET
For reproducibility of the results, the reporting of data analysis and datasets should also be standardized.Most studies lack reporting of meta-information, for example, the number of sample images within the dataset, preprocessing steps for data cleaning and data analysis.Furthermore, the literature also reports the dependence of segmentation performance on image quality and quantity.For better segmentation performance, the training algorithm needs big data of high-quality detailed images, i.e., a large number of samples per subject and a high number of subjects within the dataset.The availability of patient follow-up images could also help in training the algorithm to learn the pattern of disease progression or regression.Currently, to our knowledge, no such publicly available datasets have multiple images of a large number of patients (ideally millions, as in the UK BioBank [107] which is not free) over the years (follow-ups).
By combined efforts and taking specific actions, computer scientists, engineers, and clinicians can address the limitations mentioned above and develop segmentation tools that are clinically relevant, accurate, and robust.

V. CONCLUSION AND FUTURE DIRECTIONS
Most existing reviews on the topic of cardiovascular neurocristopathy focus on the progression and regression of disease related to abnormal development or migration of cardiac neural crest cells to detect and segment such diseases.The application of machines and DL models in cardiovascular imaging is rapidly increasing with the aim of a possible reduction in reporting time with high precision for the segmentation of cardiovascular neurocristopathy diseases, such as cardiomyopathy, stroke, aorta, and aortic valve dysfunction diseases.In this review, the available literature on the segmentation of cardiovascular neurocristopathy diseases using imaging modalities, such as CT, MRI, X-ray and Echos, is analyzed in terms of implemented segmentation algorithm, performance metrics used, meta-information about the dataset and dataset availability.
From the literature review, it can be concluded and is evident from Table 3, Table 4, Table 5 and Table 6 that Computer Tomography (CT) images are the most commonly used image modality for the segmentation of cardiovascular neurocristopathy disease, while MRI is the second best.Limited literature is available on the use of radiographic (X-ray) and echocardiography (Echo) images for the segmentation of such diseases.This is an open research topic as better imaging equipment has been developed that provides a good image that could help in more robust segmentation using these modalities.In terms of the most frequently used DL model, the segmentation using UNet, and its variants, outperform all the other models.UNet is a type of convolutional neural network (CNN) which has features of skip connection (to capture high-and low-level features of image), awareness of spatial context (deals with complex background), symmetric architecture (to capture fine-grained details) and data augmentation ability enabling them to be trained on a smaller dataset.These features make UNet better than other networks in the image segmentation task.In the literature, Simple CNN is the second most frequently implemented algorithm.While some studies have used adversarial generative networks for the image segmentation task, further investigation into the use of such models is required to get trustworthy/acceptable results.
The review also identified gaps in developing effective disease segmentation models using medical images, especially for critical conditions like cardio-neurocristopathy diseases.Limited data availability, which stems from privacy concerns and the scarcity of expert annotations, necessitates innovative solutions like data augmentation and transfer learning.Ensuring the accuracy and consistency of annotations is critical to mitigate model biases.The diverse interpatient variability in anatomical and pathological aspects underscores the importance of techniques such as data augmentation and domain adaptation to enhance model generalization.Interpretability is crucial for gaining medical professionals' trust, and the incorporation of explainable AI methods can facilitate model transparency.The transition from research to clinical practice involves rigorous validation through collaboration with medical experts and institutions while optimizing models for real-time processing is essential for timely clinical decision-making.Addressing these challenges requires the combined efforts of researchers/engineers and clinicians.Table 8 summarizes the problems and contributions of each stakeholder to overcome these shortcomings.
In the future, in conjunction with collaborative efforts, the use of intravascular ultrasound (IVUS), optical coherence tomography (OCT), and functional magnetic resonance imaging (fMRI) techniques for the detection and segmentation of cardiovascular neurocristopathy disease should also be investigated.IVUS could be used to identify the location and extent of cardiovascular abnormalities associated with neurocristopathy diseases.Brain activities, such as cognitive deficits, determined by fMRI could be associated with congenital heart disease, a type of cardiovascular neurocristopathy.OCT imaging could help visualize the microstructure of blood vessels within the heart and identify abnormalities in the cardiovascular system associated with neurocristopathy diseases.All of these techniques may not provide direct visualization of abnormalities in the cardiovascular system, but a correlation can be derived from their analysis that could help clinicians in the early detection of cardiovascular neurocristopathy diseases.
Furthermore, Meta AI has proposed an ImageBind model [108] that capitalizes on image-paired data to create a unified representation space.An example of such a model is shown in Figure 9.The ImageBind model learns joint embeddings across six different modalities which are image, audio, depth, text, thermal and Inertial Measurement Unit (IMU) data.Utilizing such a model, a medical data-based model could be designed in the future that could generate the cardiac parameters (cardiac rhythm, beat sound) and medical reports from a single image (CT, MRI, X-ray and/or Echo) using a few-shot DL approach, reducing the cost of healthcare and providing an efficient diagnosis.

FIGURE 2 .
FIGURE 2. Anatomy of different cardiovascular neurocristopathy diseases.(a) Aortic aneurysms: where a balloon-shaped bulge is formed in the aorta.(b) Aortic dissection: where there is a tear in the inner layer of the aorta.(c) Bicuspid aortic disease: where an aortic valve contains only two cusps (or flaps) instead of three.(image source: Wikimedia Commons, license: CC BY-SA 4.0, 3.0 and 4.0, respectively).

FIGURE 3 .
FIGURE 3. Overview of image segmentation tasks for different imaging modalities.Adopted from [8].

FIGURE 4 .
FIGURE 4. Adopted PRISMA guidelines and papers retrieved in this review.

FIGURE 6 .
FIGURE 6. Magnetic resonance imaging; Cardiomyopathy segmentation.(A) Images acquired in patients without cardiomyopathy (B) Images acquired in patients with Dilated Cardiomyopathy (DCM) (C) Images acquired in participants with Hypertrophic cardiomyopathy (HCM).Image Adapted from [93]: open access.

FIGURE 7 .
FIGURE 7. X-ray: Predicted mask annotated with the cardiothoracic ratio.The clavicle (collar bone) is shown in red colour, the lungs in blue and the heart is shown in green colour.Adapted from [100]: open access.

Figure 7
Figure 7 shows the segmentation of the chest anatomy to calculate the cardiothoracic ratio.(a) shows the original chest X-ray image, (b) is a segmented image using CardioNet, and (c) illustrates the calculation of the maximum width of the heart and thorax using an equation 1.

1 )
CONVOLUTIONAL NEURAL NETWORKS (CNN) Convolutional Neural Networks (CNNs) have proven their versatility in various computer vision tasks, including image segmentation.The encoder extracts hierarchical features from the input images through convolutional layers, while the decoder upscales these features to produce pixel-wise segmentation maps.Activation functions like ReLU are common, and loss functions, such as cross-entropy or Dice loss, are employed to optimize the model.Optimization techniques such as stochastic gradient descent (SGD), Adam, RMSprop, or adaptive learning rate algorithms are used for weight updates.Mini-batch training is preferred for improved convergence and memory efficiency.Data augmentation, involving random rotations, flips, and scaling, expands the training dataset and reduces overfitting.

TABLE 1 .
Comparison of recent existing review articles with our review article.

TABLE 2 .
Performance metrics used in the medical domain for segmentation assessment.

TABLE 3 .
Frequently used DL and ML algorithms in the field of cardiovascular neurocristopathy.Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.need,models can be developed that generate heatmaps or attention maps to highlight regions of interest in the images.Explainable AI techniques, such as LIME or SHAP, can provide insights into model decisions.Model Validation and Clinical Trials: Transitioning from research models to clinical practice requires rigorous validation and testing.

TABLE 4 .
Learning models for disease segmentation using Computer Tomography (CT) images.

TABLE 5 .
Learning models for disease segmentation using Magnetic Resonance Imaging (MRI).

TABLE 6 .
Learning models for disease segmentation using X-rays and Echo-cardiogram (Echo) images.

TABLE 8 .
Specific actions required by engineers and clinicians to address the limitations identified in proposed review.