Congenital heart defects (CHDs), which affect nearly 40,000 newborns annually in the United States, are the most frequent cause of infant death from birth defects [1], [2], [3]. Since the first palliative CHD corrective surgeries in the mid-20th century, great improvements have been made in patient care, noninvasive diagnosis, and surgical technique. From the 1950s up to the early 2000s, patient survival into adulthood increased from below 25% to greater than 90% [4,5]. The number of CHD surgeries is growing every year, and surgical techniques and technology are continually refined. Despite this, no significant improvement in survival has been observed during this century [6]. Although greater than 90% of these patients now survive into adulthood with few lasting complications, the patients demonstrate an increased risk of arrhythmias, heart failure, stroke, endocarditis, and sudden cardiac death [1,7,8]. Although the surgical description of “total correction” is sometimes given to these patients after successful surgery, the burden of congenital heart disease continues to be carried throughout the entire life of the patient. The methods, techniques, and residuals of initial corrective surgeries have a large impact on the patients’ long-term health and quality of life [9]. While a wide variety of lesions, scarring, and damage to cardiac tissues can contribute to post-operative complications, damage to the cardiac conduction system (CCS), in particular, carries with it the potential for severe complications such as complete heart block and sick sinus syndrome [10], [11], [12], [13].

Since the first palliative CHD corrective surgeries in the 1950s, surgeons have explored means to identify the CCS and avoid damaging it during surgery [14,15]. Although numerous approaches and technologies have been evaluated across more than half a century, the current state-of-the-art method for identifying and avoiding the CCS and its supporting vasculature during surgery continues to be the use of anatomical landmarks, which was pioneered in the 1800s and refined with modern histological techniques [16]. This practice involves using anatomical landmarks such as the cavoatrial junction, terminal crest, coronary sinus, triangle of Koch, and others to approximate the locations of the CCS components and their supporting tissue structures, such as the Sinus and AV nodal arteries. Decades of focused research have led to our current understanding of nodal tissue arrangement based on anatomical landmarks. This has enabled surgeons to leverage these anatomical landmarks to carry out corrective surgical procedures. Numerous surgeons, anatomists and researchers such as Thomas James, Jesse Edwards, Francis Fontan have established this bedrock of surgical and anatomical understanding that supports modern corrective surgeries for congenital defects [17], [18], [19], [20], [21], [22], [23], [24]. The use of anatomical landmarks during surgery provides a useful, gross estimate of the location of the nodal tissues and their supporting structures.

Studies since the 1970s revealed how the locations of the CCS with respect to superficial anatomical landmarks are more variable in congenitally deformed hearts than in normal hearts [25], [26], [27]. Recent analyses have shown how, in addition to the sinus and AV nodal tissues, the location of the nodal arteries and other critical nearby structures like nervous tissue, which are also not visible to the naked eye, can vary significantly with respect to the anatomical landmarks, even in normal human hearts [28], [29], [30]. As more light is continued to be shed on nodal anatomy variability, the regions where CCS components might be present increase in area, subsequently increasing the size of “danger zones” in the cardiac tissue. The caution surgeons take when avoiding the cardiac tissue danger zones is reflected in an increased number of residual lesions left after the surgery, which negatively impacts postoperative patient health [15,31,32]. As is evidenced by the continued prevalence of heart block, postoperative arrhythmias, ventricular desynchrony, and eventual heart failure, the use of anatomical landmarks alone to identify the CCS is insufficient to prevent iatrogenic damage [8,10,33,34]. Although much research has been done to identify and describe the locations and variations of the cardiac conduction system components, the number of postoperative permanent pacemaker implantations has continually increased in the decades since the 1960s.

Damage to the CCS during corrective surgeries can occur due to unexpected anatomical variations of the CCS [10]. To better understand and communicate the dispositions of the CCS components, studies of the nodal tissue regions have used a variety of techniques, including observations and 3D reconstruction from serial histological tissue sections, contrast-enhanced micro-computed tomography (CT), and magnetic resonance microscopy (MR) [35], [36], [37]. While these studies have provided actionable insights for surgeons, and constitute the basis for current surgical practice, these studies typically focus on gross anatomical structure. The finer details of the CCS have continually evaded high-resolution 3D representation and visualization, obscuring anatomical variations and interactions between the nodal tissues and the surrounding tissue architecture [28,29]. The CCS remains challenging to visualize at high resolutions because of its small size, depth beneath the tissue surface, lack of CCS-specific biomarkers, and difficulty producing and processing large datasets of serial histological sections [38]. Because of these limitations, visualizations created of the nodal tissue structures lack information on the relationships between the nodes and their crucial supporting vasculature.

Within the last five years, advances in the field of machine learning and advanced computer vision methods have had a major impact on the medical imaging field [39]. Large high-performance compute clusters, and deep learning computational approaches coupled with whole-slide imaging have enabled greater diagnostic and image processing capabilities. These methods are now beginning to surpass the capabilities of expert human visual inspection [40].

This study aims to overcome the limitations of past approaches when studying the 3D micro-anatomy of the cardiac conduction system to produce high-resolution, data-rich models that elucidate anatomical variations in normal neonatal human hearts. To do this, we leveraged modern image processing and machine learning methods to create 3D models of the sinus and AV nodal regions. Our approach encompasses the use of a modified U-Net deep convolution neural network architecture to enable the high-throughput segmentation of near whole-slide serial images extracted from human pediatric nodal tissue samples. These segmentations, which contain information on numerous tissue types including nodal tissue, vasculature, and neural tissues, are then used to produce high-resolution 3D visualizations as well as perform quantifications of tissue characteristics.

By leveraging modern high-performance computing methods, this approach enables the production of high-resolution data-rich 3D models as well as numerous novel visualizations of the tissues of interest, adding precise insight into the disposition and relationship of the nodal tissues with respect to their surrounding structures. These models describe novel variability in the sinus and AV nodal regions, with implications for guiding future surgical strategies that could potentially decrease iatrogenic damage to the cardiac conduction system while enabling a more concrete and precise understanding of inherent anatomical variability of the heart.