Advancements in Artificial Intelligence for Fetal Neurosonography: A Comprehensive Review

1. Introduction

The Assessment of the anatomic integrity of the fetal central nervous system (CNS) is one of the most challenging tasks during prenatal sonographic work-up as brain development and maturation constitutes a complex and well-orchestrated process at various embryonic and fetal stages. To preclude diagnostic errors, national and international guidelines explicitly paid attention to the fact that the appearance of the brain underlies profound changes throughout gestation. Although brain anomalies are among of the most common fetal malformations with an estimated prevalence of 9.8-14 per 10.000 live births [1,2], their in utero perception fundamentally requires familiarity with sonographic brain anatomy, artifacts and a designated vigilance for the necessity of subsequent targeted multiplanar assessment of the entire fetal CNS (neurosonography) [3,4]. In general, the efficacy of ultrasound (US) screenings largely hinges on the operator’s skill in navigating to and reproducing standard imaging planes and this, in turn strongly relates to the gestational age (GA) at examination. In this context, it could be noticed that in the recent past the majority of severe congenital brain anomalies were readily identified prenatally by applying a systematic, protocol-based US survey [5,6]. Nevertheless, detection rates of fetal brain lesions in an unselected population remain somewhat unsatisfactory and more subtle changes might escape an early diagnosis. In part, this might be explained by the fact that even though advanced technologies as three-dimensional US (3DUS) undoubtedly have the potential to contribute to an improved detailed CNS evaluation, there is still little consensus as to the ideal method of volume acquisition, settings, and analysis of the volume and an overall lack in standardization of volumetric assessment [7]. On the other hand, DiMascio et al. stated in their systematic review that fetal brain charts substantially suffer from poor methodology and are at high risk of biases, mostly when focusing on relevant neurosonographic issues [8]. In addition, another publication demonstrated that fetal cortical brain development in fetuses conceived by assisted reproductive technology seems to be different from those conceived spontaneously, as expressed by a reduced sulci depth [9]. This underpins the complexity of an all-encompassing thorough assessment of the fetal brain. Beyond any doubt, prenatal US is capable of providing precise information regarding fetal anatomical integrity and the severity of abnormal conditions derived from high-quality images with increased diagnostic accuracy and reliability. The transabdominal route remains the technique of choice for comprehensive anatomic evaluation of dedicated organs like the fetal brain. This clearly demonstrates that the currently available image data source has to deal with a combination of maternal, fetal, technical, environmental, and acoustic factors hampering image clarity, data acquisition and eventually establishing precise antenatal diagnoses.

Current research approaches regarding the clinical applicability of artificial intelligence (AI)-assisted methods in the context of fetal neurosonography (beyond the first trimester) are heterogeneous and, with few exceptions, software solutions being of use for clinical routine are rather rare. However, several promising research topics in this field has emerged. These mainly include (among others) optimized (automated) acquisition of 2D standard planes with correct orientation and localization within a 3DUS volume, a simplified workflow, automated recognition of crucial CNS and bony structures (as landmarks) with subsequent detection of anomalies, evaluation of image quality and assessment of GA by evaluating neurodevelopmental maturation [10,11,12,13,14].

This review aims to summarize current knowledge about potential diagnostic targets for AI algorithms in assessment of the fetal brain in a clinical context and highlights why AI applications are increasingly being integrated into prenatal US interrogation and illuminate their practical added value.

2. AI in Prenatal Diagnosis

The In the very recent past, we are witnessing the tidal wave of artificial intelligence and their computational applications in healthcare in general and in medical image analysis in particular. In 2022, Dhombres et al. published a systematic review regarding the actual contributions of AI reported in obstetrics and gynecology (OB/GYN) discipline journals. In detail, most articles covered method/algorithm development (53%, 35/66), hypothesis generation (42%, 28/66), or software development (3%, 2/66). Validation was performed on one data set (86%, 57/66), while no external validation was reported [15].

Machine learning (ML) is a powerful set of computational tools that learn from large (structured) data sets and trains models on descriptive patterns and subsequently applies the knowledge acquired to solve the same task in new situations. Although ML algorithms are presently widely deployed in medicine expanding diagnostic and clinical tools to augment iterative, time-consuming and resource-intensive processes and streamline workflows, considerable human supervision is needed. AI models that use Deep learning architectures (DL; a subdomain of ML) which predominantly leverage large-scale neural networks mimicking silicon circuit synapses tend to outperform traditional machine learning methods in complex tasks and constitute the most suitable methodology for image analysis. Based on a detailed scoping review dealing with the most-cited papers using DL in literature from 2015 to 2021, the number of surveyed publications for segmentation, detection, classification, registration, and characterization tasks comprised 30, 20, 30, 10, and 10 percent, respectively [16]. It is of note that the quality of obstetric US screening images is crucial for clinical downstream tasks, involving assessment of fetal growth and development, in utero compromising, prediction of preterm birth and detection of fetal anomalies, respectively. It is now widely recognized by leading US equipment manufacturers and most of the experts in this field, that there are clear benefits of utilizing AI technologies for US imaging in prenatal diagnostics. A multitude of Convolutional Neural Network (CNN)-based AI applications in US imaging have showcased that AI models can achieve comparable performance to clinicians in obtaining appropriate diagnostic image planes, apply appropriate fetal biometric measurements, and accurately assessing abnormal fetal conditions [10-13,17,18].

As accurate head measurements are of crucial importance in prenatal and obstetrical ultrasound surveillance, a plethora of automatic methods for fetal head analysis have been proposed. Most studies focused preferentially on size and shape of the bony skull - excluding internal structures - while applying solely head detection methods (e. g. object (skull) detection using bounding boxes, segmentation methods ± ellipse fitting, edge-based and contour-based methods). Torres et al. published an excellent comprehensive state-of-the-art review tabulating more than 100 published papers on computational methods for fetal head, brain and standard plane analysis in US images [13]. Moreover, the survey also summarized image enhancement protocols of US images, including methods that find the fetal head aligned to a coordinate system, compounding approaches, as well as US and multimodal registration methods. The authors provided an exhaustive analysis of each method based on its clinical application and theoretical approach and in their concluding remarks they stated - despite the fact that a multitude of distinct image processing methods has been developed in the recent past (which mainly comprised deep learning approaches) - there is a need for new architectures to boost the performance of these methods. A strong database is seen as an indispensable prerequisite, which reinforces the need for public US benchmarks and for the development of approaches that deal with limited data (e.g., transfer learning approaches). And on the other hand, more effort should be made in AI research for the development of methods to segment the head in 3D images as well as methods to reliable detect abnormalities or (even subtle) lesions within the fetal brain. Accordingly, Ramirez Zegarra and Ghi suggested an ideal AI setting which clearly addresses the urgent need for more multitasking DL models that are trained for both the detection of fetal standard planes, identification of fetal anatomical structures and performance of automatic measurements that in turn will consequently be able to generate alarm messages in the event of malformations [19].

In fact, in the last decade several AI-related scientific studies have been conducted to improve the quality of prenatal diagnoses by focusing on three major issues: (I) detection of anomalies, fetal measurements, scanning planes, and heartbeat, (II) segmentation of fetal anatomic structures in US still frames and videos and (III) classification of fetal standard diagnostic planes, congenital anomalies, biometric measures, and fetal facial expressions [20].

Various researchers were able to develop algorithms that were able to reproducibly quantify biometric parameters with high accuracy. Some of them will be discussed critically below. On the other hand, quite a number of AI models were trained with inadequate and/or insufficiently labeled samples, which led to overfitting problems and performance degradation [14]. ‘All models are wrong, but some are useful’, an aphorism on the subject of statistics more than 50 years ago coined by George E.P. Box, that describes the general dilemma in the targeted application of computational modeling approaches and AI solutions (not only in the past) [21].3. Results

This section may be divided by subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.

3. AI in Fetal Neurosonography

Adopting experiences made with automated techniques in fetal cardiac assessment, further refinements in AI algorithms or development of anomaly-specific learning algorithms could help achieve more granular detections of unique CNS lesions [22]. This has the potential to risk stratify certain fetal populations. But, as a matter of fact, it has to be acknowledged that algorithms developed for fetal imaging recognition require a larger database compared with other AI algorithms, due to the similar appearance of the different US planes [19]. However, it should be noted that currently available and clinically approved approaches for the extensive processing of three-dimensional data sets of the fetal CNS do only sparsely exploit the diagnostic potential of volume US. In fact, it is currently not feasible to perform both simple and more complex tasks such as assessment of the total brain volume, rendering of the brain surface, slicing and display of all diagnostic sectional planes (in accordance with the ISUOG guidelines) in addition to multiplanar image display on one and the same 3D volume (vendor-independent). This applies to both conventional tools and AI frameworks and does not appear to be readily explainable in the light of the existing scientific literature with generally highly complex AI pipelines. In this regard, developers and engineers on the one hand and clinicians on the other should work together more intensively to find relevant integrative volume-based solutions for clinical routine as quickly as possible.

In 2023 an international research group developed a normative digital atlas of fetal brain maturation based on a prospective international cohort (INTERGROWTH-21st Project) using more than 2,500 serially acquired 3D fetal brain volumes [23]. In preparation of this fully-functional digital brain atlas, the authors proposed an end-to-end, multi-task CNN that both extracts and aligns the fetal brain from original 3D US scans with a high degree of accuracy and reliability (Brain Extraction and Alignment Network; BEAN) [24,25,26]. These obligate steps were mandatory (as in most neuroimage analysis pipelines) to enhance the visibility of brain structures within the 3D US templates and on the other hand to significantly reduce the amount of processed extra-cranial volume information and lastly to overcome the positional variation of the brain inside the scan volume. From the author’s perspective, there is no doubt that introduction of computerized human body atlases either based on US or MRI image data (as published earlier by Gholinpour et al. [27]) will contribute to our understanding of fetal developmental processes in general and brain maturation in particular by providing rich contextual information of the inherently 3D (CNS) anatomy.

The clinical applicability of semiautomatic volumetric approaches in terms of a detailed reconstruction of diagnostic planes of the fetal brain has been validated in previous studies [28,29,30]. Very recently a 3D UNet-based network for 3D segmentation of the entire CNS using an intelligent navigation to locate CNS planes within the 3D volume has been introduced (fully automated 5DCNS+™). While applying this tool, our group was able to show that CNS volume data sets (acquisition originated from an axial transthalamic plane) could readily be reconstructed into a nine-view template in less than 12 seconds on average facilitating the generation of a complete neurosonogram with high accuracy, efficiency and reduced operator-dependency, confirming previous findings, accordingly (data not shown).

Lu et al. reported on an automated software (Smart ICV™) to total fetal brain volume retrieved from 3DUS volume data. This novel technique showed a high intra- as well as inter-observer intra-class coefficient (0.996 and 0.995, respectively) and high degree of reliability compared with a manual approach using Virtual Organ Computer-aided AnaLysis (VOCAL™) [31]. An overview of current Ai-driven algorithms with either clinical or pre-clinical context is given in Table 1.

DL algorithms have become the methodology of choice for imaging analysis [16,18,32,33]. DL models are capable of overcoming US image-related challenges including inhomogeneities, (shadowing) artifacts, poor contrast and intra- and inter-clinician acquisition and measurement variability, respectively. Fiorentino and co-workers categorized published work in the field of fetal US image analysis applying a plethora of different DL algorithms. Their review surveyed more than 150 research papers to elaborate the most investigated tasks addressed using DL in this field [18]. The authors could demonstrate that fetal standard-plane (SP) detection (19.6%) and fetal-biometry estimation (20.9%) were among the most prevalent tasks. Fetal CNS and heart were the most explored structures in standard plane detection while fetal head circumference was the most frequently investigated measurement in biometry estimation. In 49 % researchers trained DL pipelines for anatomical structure analysis. The most studied anatomical structures were heart and brain, contributing with 26.7% and 20.0% of the surveyed papers, respectively. In case of the latter, the analysis is performed both with 2D and, more recently, 3D images to assess brain development and localization, structure segmentation and GA estimation. Challenges to be addressed regarding AI in image analysis in general comprise: limited availability of (multi-expert) image annotation, limited robustness of DL algorithms due to the lack of large training data sets (interestingly only a minority of DL studies use data from routine clinical care), inconsistent use of both performance metrics and testing datasets hampering a fair comparison between different algorithms and research proposing semi-, weakly or self-supervised approaches is still scarce [10,18,32].

3.1. AI in GA Prediction

Reliable methods for accurate GA estimation in the second and third trimester of pregnancy remains an unsolved challenge in obstetrics. This might be addressed to late booking, infrequent access to prenatal care, and unavailability of early US examination and other reasons [63]. Namburete et al. introduced a model which was able to characterize the neuroanatomical appearance both spatially and temporally while identifying relevant brain regions such as the Sylvian fissure, cingulate and callosal sulci as powerful image regions in the GA-discrimination task [42]. The authors extended in addition to clinically relevant metadata like the head circumference canonical features sets (e.g., Haar-like features) to capture structural changes within the fetal brain. The algorithm improved the confidence of age predictions provided by the clinical HC method by ±0.64 days and ±4.57 days in the second and third trimesters, respectively. A similar approach estimates GA on standard transthalamic axial plane images using a supervised DL model (quantusGA) that automatically detects the position and orientation of the fetal brain by detecting the skull and five internal key points (necessary to crop and rotate the brain, resulting in a horizontally aligned brain image). The model then extracts textural and size information from the brain pixels and uses this information to give an estimate on the respective GA with a similar or even lower error compared with fetal biometric parameters, especially in the third trimester [63]. AI models are capable of estimation of GA with an accuracy comparable to that of trained sonographers conducting standard fetal biometry (e. g. fetal head) as the results of a recent study suggest. The authors trained an DL algorithm to estimate GA from blind US sweeps and showed that the model’s performance appears to extend to blind sweeps collected by untrained providers in low-resource settings [84]. Similar results were demonstrated by two groups where ML-based algorithms outperformed current ultrasound-based clinical biometry in GA prediction with a mean absolute error of 3.0 and 4.3 days [85] or 1.51 days (using an ensemble model of both unlabeled image and video data) in second and third trimester fetuses [86].

3.2. AI for Augmenting Fetal Pose Estimation and CNS Anomaly Assessment

In the recent past, CNNs and other deep learning architectures were trained to recognize and predict fetal poses from imaging data [25,66,87,88]. In contrast to already established methods, which were mainly designed for standard plane identification assuming that a good US image quality has already been achieved with the fetus in a proper position and therefore only assist in prenatal image analysis, a study group from the UK recently emphasized the utility of recognizing the probe’s proximity to diagnostic CNS planes, facilitating earlier and more precise adjustments during 2D US scanning. The semi-supervised segmentation and classification model used an 18-layer residual CNN (ResNet-18) that was trained on both labeled standard planes and unlabeled 3D US volume slices to filter out frames lacking the brain and that was able to generate masks for those containing it, enhancing the relevance of plane pose regression in a clinical setting [81]. In a previous study the authors applied a similar 18-layer residual CNN as a backbone for feature extraction (with the pre-trained ImageNet weights) and 6D pose prediction (refers to the task of determining the six degree-of-freedom pose of an object in 3D space) of arbitrarily oriented planes slicing the fetal brain US volume without the need for real ground truth data in real-time or 3D volume scans of the fetus beforehand [72].

Yeung and colleagues proposed an algorithm for a more general task of predicting the location of any arbitrary 2D US plane of the fetal brain in a pre-defined 3D space [66]. In their work they demonstrated that, based on extensive data augmentation and complementary information from training volumes acquired at different orientations, the prediction made by a novel CNN model generalizes to real 2D US acquisitions and videos, despite the model having only been trained with artificially sampled 2D slices. Considering that 3D volumes provide more effective spatial information and exhibit higher degrees of freedom (DoF), increased variations in fetal poses make proper learning for these algorithms challenging. In fact, 6D fetal pose estimation refers to the process of determining the six degrees of freedom (6DoF) position including three translational (position) and three rotational (orientation) parameters, allowing for a comprehensive understanding of the fetus’s spatial position in a coordinate system and movement in utero. The study conducted by Chen and co-workers dealt with a similar topic - fetal pose estimation in 3D US. The authors introduced a novel 3D fetal pose estimation framework (FetusMapV2) which was able to identify a set of 22 anatomical landmarks of first and early second trimester fetuses and their specific connections to provide a comprehensive and systematic representation of the fetal pose in 3D space, to overcome challenging issues such as poor image quality, limited GPU memory for tackling high dimensional data, symmetrical or ambiguous anatomical structures, and considerable variations in fetal poses [87].

Xu and co-workers trained a DL algorithm (a cycle-consistent adversarial network (Cycle-GAN) to simulate realistic fetal neurosonography images specifically generating third-trimester pregnancy US images from second-trimester images that were qualitatively evaluated by experienced sonographers [56]. The vast majority (84.2%) of the simulated third-trimester images could not be distinguished from real third-trimester images in this study. These generative adversarial networks (GANs), first introduced by Goodfellow et al. in 2014, are algorithmic architectures that use two neural networks, competing against the other and learn to generate new, synthetic instances of data with a probalistic model that can pass for real data/images, augmenting existing datasets for training DL models [89,90]. Generative approaches can better handle missing data in multi-modal datasets by generating the missing image information and preserving sample size, thereby boosting downstream classification performance [91,92]. GANs might also assist in analyzing abnormal fetal anatomical structures (e. g. CNS anomalies) while considering also the corresponding GA information (wide range of physiological changes among trimesters leading to marked inter- and intra-organ variability) [18]. A recent research paper introduced a state-of-the-art framework (FetalBrainAwareNet) that leverages an image-to-image translation algorithm and utilizes class activation maps (CAM) as a prior in its conditional adversarial training process being capable of producing more realistic synthetic images, resulting, according to the authors, in a greater clinical relevance than similar experimental approaches [56,67,93,94]. The uniqueness of this approach was the incorporation of anatomy-aware regularization terms - one ensuring the generation of elliptical fetal skulls while the other was crucial for refining and distinctly differentiating key anatomical landmarks (e.g., cerebellum, thalami, cavum septi pellucidi, lateral ventricles) in each particular fetal head standard plane (FHSP).

In US imaging the presence of speckle noise degrades the signal-to-noise of US images, traditional image denoising algorithms often fail to fully reduce speckle noise and retain image features. A recently proposed GAN based on U-Net with residual dense connectivity (GAN-RW) achieved the most advanced despeckling performance on US images (e. g. fetal head) in terms of the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and subjective visual effect [95]. Yeung et al. proposed a novel framework (ImplicitVol), a sensor-free approach to reconstruct 3D US volumes from a sparse set of 2D images with deep implicit representation. The authors stated that their algorithm outperformed conventional approaches in terms of image quality of the reconstructed template, as well as the refinement of the spatial 3D localization, which underscored the additional potential in slice-to-volume registration [82]. The latter refers to vital technique in medical imaging that transforms 2D slices into a cohesive 3D volume, thereby enhancing the ability to visualize and analyze complex anatomical structures, leading to optimized diagnostic (and therapeutic) outcomes. The same group introduced a multilayer perceptron network (RapidVol) to speed up slice-to-volume ultrasound reconstruction following a tri-planar decomposition of original 3D brain volumes and were able to demonstrate a threefold quicker and a 46 % more accurate complete 3D reconstruction of the fetal brain (collected as part of the INTERGROWTH-21st study) compared to the aforementioned implicit approach [96].

Lin et al. developed a real-time artificial intelligence-aided image recognition system based on the YOLO (You Only Look Once) algorithm (Prenatal Ultrasound Diagnosis Artificial Intelligence Conduct System; PAICS) being capable of detecting a set of fetal intracranial malformations. The algorithm was trained on 44,000 images and 169 videos and achieved an excellent performance on both internal and external validation, with accuracy comparable to that of expert sonologists [73]. The same group conducted a randomized control trial that assessed the efficacy of a deep learning system (PAICS), in assisting fetal intracranial malformation detection. More than 700 images/videos were interactively assessed by 36 operators with different levels of expertise. With the use of PAICS (prior or after individual interpretation) an increase in detection rates of fetal intracranial malformations from neurosonographic data could be noticed [77].

3.3. Other Current AI Applications Related to Fetal Neurosonography

The recent rapid emerging subfield of AI concerning the interaction between computers and human language is known as natural language processing (NLP). The launch of the chatbot ChatGPT, a large language model (LLM), in 2022, based on an NLP model known as the Generative Pretrained Transformer (GPT), has generated a wide range of thinkable applications in healthcare [97,98,99]. Therefore, beyond the field of (often cited) scientific writing, identifying suitable areas of application in obstetrics and gynecology, including fetal neurosonography, is obvious. It is crucial to understand that ChatGPT, trained on massive amounts of text data, mimic statistical patterns of human language, generates outputs based on probabilities and thus, emulating human conversation dynamics [97,98,100,101]. Before addressing potential applications of ChatGPT in the context of fetal neurosonography, two fundamental, its capability limitating aspects, must be kept in mind, whose knowledge must never be ignored in the interpretation of the subsequent thoughts: although ChatGPT should increasingly be capable of generating meaning-semblant behavior, the current technology lacks semantic understanding [102]. Be aware of hallucinations and fact fabrications [98]. Furthermore, the generated content suffers from an absence of verifiable references [99,100,101].

Most recently, the latest version of ChatGPT (GPT-4) has been evaluated for its use to facilitate referrals for fetal echocardiography to improve early detection and outcomes related to congenital heart defects (CHD) [103]. Kopylov et al. found moderate agreement between ChatGPT and experts. Comparing AI referrals to experts indicated agreement of around 80% (p< 0.001). For minor CHD cases, the AI referral rate was 65% compared to 47% for experts. In future, AI could presumably support clinicians in this area.

A similar approach would be conceivable for fetal neurosonography, as well as the implementation of language-based AI support in summarizing the findings and optimizing the wording of complex medical reports or even the classification of various sonographic CNS abnormalities into corresponding disease entities with differential diagnoses. In summary, ChatGPT cannot be used independently of experts in the field of fetal neurosonography and certainly will not replace them [100,104]. We agree with other authors concerning its unlikely that ChatGPT, even in improved versions, will ever be able to provide reliable data at the standard required by evidence-based medicine [100,105]. However, repetitive, time-consuming tasks and trains of thought can soon be left to this chatbot in clinical routine.:

4. Perspectives

AI-based applications, on whose algorithms prenatal diagnostics will increasingly depend on, are fundamentally changing the way clinicians use US. Even if the development of AI-based applications in obstetric US is still in its infancy and automation has not yet reached the required level of clinical application, any time soon, the use of AI in fetal neurosonography would exceed the capabilities of human experts, as in other fields of fetal US (11,14).

A recently published paper introduced a novel approach that parameterizes 3D volume data as a deep neural network, which jointly refines the 2D-to-3D registrations and learns a full 3D reconstruction based on only a set of non-sensor-tracked freehand 2D scans [82].

Unfortunately, the black-box nature of most machine learning models remains unresolved, and many decisions of intelligent systems still lack interpretable explanation. Explainable AI (XAI) is deemed to provide methods, equations and tools to make the results generated by an AI algorithm comprehensible for the user. By providing visual and feature-based explanations, XAI enhances the transparency and trustworthiness of AI predictions and could thus pave the way of initial uptake of an AI model into clinical routine [106]. In this regard, a recent study analyzed the performance of several CNN trained on 12,400 images on fetal (CNS) plane detection after input (image) enhancement by adopting Histogram Equalization and Fuzzy Logic based contrast enhancement. The results achieved an accuracy between 83.4 and 90 % (depending on the classifier analyzed) and were subsequently evaluated by applying LIME (Local Interpretable Model-Agnostic Explanations) and GradCAM (Gradient-weighted Class Activation Mapping) algorithms to examine the decision-making process of the classifiers, providing explainability for their outputs [107]. These XAI models visually depict the region of the image contributing to a particular class, thereby justifying why the model predicted that class [108]. Very recently, Pegios and co-workers used iterative counterfactual explanations to generate realistic high-quality CNS standard planes from low-quality non-standard ones. With their experimental approach (Diff-ICE) they demonstrated the superior value for the challenging task of fetal ultrasound quality assessment as well as its potential for future applications [109]. To alleviate the risks of incomprehensibility and - more crucial - clinical irrelevance in forthcoming research two publications proposed directive guidelines for transparent ML systems in medical image analysis (INTRPRT/Clinical XAI Guidelines) [110,111]. Interestingly, all sixteen commonly used heatmap XAI techniques evaluated by Jin et al. were found to be insufficient for clinical use due to their failure in the criteria ‘truthfulness’ and ‘plausibility’ [111].

Acknowledging the recent achievements of AI in medical image analysis, Sendra-Balcells and co-workers addressed the paradox that the development of AI in rural areas in the world, like the Sub-Saharan Africa, is at its lowest level and on the other hand the current AI advancements including deep learning implementations into prenatal US diagnosis that facilitate an improved antenatal screening. In this regard they investigated the generalizability of fetal US deep learning models to low-resource imaging settings [112]. The authors pre-trained a DL framework for standard plane detection (e. g. fetal brain) in centers with greater access to large clinical imaging datasets and subsequently applied this model to African settings. These results gained from transfer learning exemplify that domain adaptation might be a solution to support prenatal care in low-income countries.

As a recent commentary given by Tonni and Grisolia correctly stated, we will inevitably face that the incorporation of AI solutions within the US apparatus will start to surge exponentially in the near future, producing beneficial effects not only in terms of diagnostic accuracy but also producing positive effects on the quality of fetal examination in its entirety, including appropriate surveying of complex anatomical structures (e. g. fetal CNS and heart), reporting, and improving medical-legal issues for physicians involved in both fetal imaging and fetomaternal care [113].

5. Conclusions

AI has increasingly been accepted as a fundamental component in a multitude of healthcare applications such as medical image analysis. In the light of this inevitable and intriguing flooding of intelligent algorithms into modern US diagnostics nothing less than the beginning of a new era of 5D ultrasound has been proclaimed. However, there are several challenges of AI deployment particularly in fetal neurosonography that must be solved: the need for large and diverse training datasets (2D/3D) in general; the difficulty of training accurate models for diagnosing evolving fetal brain abnormalities, and the potential for algorithmic biases; the urgent need to address the troubling lack of transparency and interpretability of current AI algorithms to achieve further translation into clinical diagnostic circuits and to avoid reluctance to use if AI models seemingly only demonstrate a benefit on the optimal patient; and the seamless integration of AI models into diagnostic workflows requires careful consideration of ethical and legal implications, as well as the need for rigorous validation studies to ensure the safety and efficacy of AI applications.

However, it remains to be seen how fast and in what manner promising techniques like 6D fetal pose estimation, slice-to-volume registration tools, and real-time recognition of normal and abnormal CNS anatomy, to name a few, will be integrated into clinical practice and medical education, alongside the continued advancement of current, already commercialized AI frameworks.