Thoughts on the contribution of artificial intelligence ( <scp>AI</scp> ) to assessment of the fetal heart: a true scientific odyssey

Abstract

The July 2024 issue of the Journal, including both an Editorial by Drukker1 and two original studies2, 3, highlights the growing excitement around the contribution of artificial intelligence (AI) in the field of fetal heart research. The development of computational science technologies has led in recent years to an exponential increase in scientific literature on the role of AI, both in prenatal medical imaging in general and in fetal heart assessment in particular. Numerous publications range from theoretical perspectives4-6 and basic preliminary results7, 8 to more recent articles addressing the particulars of ultrasonographic screening9, 10 and diagnosis2, 3, 11-13 of congenital heart disease (CHD) during pregnancy. In this Opinion, we discuss some considerations that have been enriched in recent years by scientific data, conference contributions and the emergence of certain solutions proposed by ultrasound manufacturers and companies for the prenatal analysis of this unique organ, the fetal heart, which still evokes both anxiety and fascination among healthcare professionals performing antenatal ultrasound examinations. When considering strategies for the use of different AI models in fetal cardiac assessment, it is important first to reflect on current practice and the causes of non-detection of prenatal CHD. In our view, AI-assisted prenatal assessment of the fetal heart needs to be approached from two different perspectives, screening and diagnostic, which have completely different constraints and goals. For both, we believe it is important to have human input to define clearly in which situations AI should be used and in which radiomics. The efficiency of CHD screening can be affected by several well-documented factors. These include maternal factors (abdominal wall thickness, presence of leiomyomas and of abdominal scars), fetal factors (gestational age at the time of ultrasound, fetal dorsal position, amniotic fluid volume and type of CHD) and operator-related factors (number of years of experience, number of ultrasound examinations performed annually and knowledge of CHD). As maternal characteristics cannot be altered and fetal positioning during the examination can only be considered in a limited and unsystematic way (because some operators prefer to move the baby while others do not), the main causes of non-detection of antenatal CHD that can be addressed are operator-related, and may be divided into three broad categories14, 15: (1) lack of knowledge of or failure to follow recommended ultrasound practice (guidelines), resulting in omission of the assessment of one or more key views (failure to analyze views of interest); (2) lack of skills in obtaining views of interest, resulting in poor-quality acquisition for interpretation; (3) inability to recognize abnormal features despite obtaining informative views of interest (inability to recognize unusual features). Unlike Achilles, who was unaware of his vulnerability, professionals performing ultrasound examinations in obstetrics must learn to recognize and accept their human limitations – technical, intellectual and in terms of concentration and alertness. Efforts should focus on developing solutions and tools that address the three aforementioned categories, preferably in real time during the ultrasound examination, with the aim of optimizing operator skills and performance, rather than replacing the operator. Theoretically, screening for CHD must have very high specificity. In practice, screening essentially comprises the large-scale implementation of recommendations that are updated regularly by various scientific societies, including the International Society of Ultrasound in Obstetrics and Gynecology (ISUOG)16. These recommend analysis that involves almost simultaneous acquisition and interpretation of specific views of interest. Assessment of these views considers both their presence (acquired or not) and whether they meet specific quality criteria. While there is consensus on which views to analyze to enable prenatal depiction of most detectable CHDs (abdominal situs, four-chamber view, left and right outflow tracts and three-vessel view), there is not yet any consensus on the associated quality criteria. However, ISUOG has recently opened the door to address this, by proposing a checklist for fetal heart analysis in the screening context16. This philosophy that prenatal ultrasonographic screening for CHD should focus on both completeness and quality of cardiac views aligns perfectly with the general principles of high-quality screening in the setting of supervised or semi-supervised AI, which requires the availability of a large amount of images for building high-quality datasets specific for screening purposes, diversity of the images in terms of patient populations, operators and devices (different manufacturers and models), the creation of repetitive analysis methods or processes, and quality assurance assessment. Several recent studies have focused on developing solutions using AI in the field of prenatal ultrasonographic screening, with articles published on automated detection of certain key views of the fetal heart5, 17, quality assessment of one7, 8, 18-24 or more24, 25 of these views, and analysis of specific structures within these views, such as the interventricular septum26. As pointed out by Sklansky and DeVore27, Yagel and Moon-Grady14, Van Nisselrooij et al.28 and ourselves15, practitioners often lack knowledge regarding how to improve image quality and are not sufficiently critical of the quality of their examinations and the insonation angles used in obtaining the required views. AI algorithms capable of recognizing the key views and assessing their quality should theoretically allow detection of most CHDs, and their failure to recognize particular expected information should lead to referral for an expert examination. As with human-based analysis, AI algorithms developed for screening must have a low false-positive rate to avoid unnecessary follow-up testing, which can be costly, stressful and burdensome for both healthcare providers and patients. Just as the Spartan armies were known for their organization and efficiency even when performing complex tasks, the philosophy behind AI solutions for ultrasonographic screening of the fetal heart must combine high specificity, robust quality control and seamless integration into operational workflows and treatment pathways, as advocated by Mahmood et al.29. Unlike CHD ultrasonographic screening, which has relatively simple goals (capturing quality views of interest and recognizing quality and normality criteria), CHD diagnosis cannot be approached with simplicity. As several teams have noted, there is some discrepancy between pre- and postnatal diagnoses, even among experts30-32. CHD diagnosis can sometimes be made solely by identifying unusual anatomical features on a single view on two-dimensional (2D) imaging (e.g. identification of a right aortic arch using the three-vessel-and-trachea view), whereas other cases may require a combined examination of the inflow and outflow tracts in both 2D and color Doppler modes (e.g. tetralogy of Fallot with pulmonary atresia and major aortopulmonary collateral arteries). Furthermore, the diagnosis of CHD involves not only classifying a fetal cardiac defect, but also incorporating additional and crucial parameters, such as: (1) performing a detailed anatomical survey to search for associated anomalies; (2) offering genetic counseling and performing prenatal exome sequencing; (3) determining the best place for delivery; (4) discussing neonatal management, including possible postnatal treatment options (e.g. prostaglandins, catheterization, surgery or hybrid treatments), and the quality of life and mid- to long-term prognosis for a child with CHD; and (5) discussing the uncertainties related to neonatal hemodynamics. Clearly, there are many challenges associated with solving the multifaceted problem of prenatal CHD diagnosis, and there is not just one solution. In contrast to the datasets available for CHD screening, which contain large quantities of images of varying quality illustrating the different views of interest, the construction of databases useful for the training of AI models for prenatal CHD diagnosis is hampered by the nature of the malformations themselves: relative rarity and diversity. It is also important to recognize that what is in these neural networks is initially constructed and monitored by humans who have imagined one or more modalities to identify unusual or pathological situations. Some teams have focused on specific pathologies, such as hypoplastic left heart syndrome12, atrioventricular septal defect2, aortic coarctation3 and ventricular septal defect25, while others have proposed a more comprehensive approach10, 33. Given the challenges in building enriched pathological datasets, several teams have questioned the relevance of the datasets used for the training of AI models in recently published studies34, many of which are images from tertiary expert centers. Questions have been raised regarding the quality (which requires diversity and heterogeneity) of these datasets. As Drukker notes in his Editorial1, the Medical AI Data for All (MAIDA) initiative35 aims to address this issue by promoting the global sharing of imaging data so that algorithms can work safely and effectively in diverse patient populations and clinical contexts. Having outlined all these limitations and challenges, prenatal CHD diagnosis can indeed be considered a true quest for the Delphic oracle, or the Holy Grail, as suggested by Drukker1. The few publications linking AI concepts to the prenatal identification of cardiac defects undoubtedly raise high hopes among readers of all disciplines. However, careful reading of these studies often leads to disappointment and the realization that there is still a long way to go. Nevertheless, solutions are beginning to emerge. These are being proposed by independent research laboratories, manufacturers and companies. The next big step will undoubtedly be to conduct high-quality, collaborative, non-retrospective studies (with heterogeneous populations), in real-world settings34. Translating existing solutions into routine care and investigating the potential impact on care pathways, clinicians and patients will play an important role in the next step towards AI solutions. Hopefully, unlike Odysseus, this scientific journey will not take 10 years to complete. E.Q. is chief medical officer of Diagnoly. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.

Ähnliche Arbeiten