Page 251 - v11i4

International Journal of Bioprinting Deep learning-based 3D digital model of fetal heart

1. Introduction for image segmentation pose significant limitations
to the clinical application of 3D printing techniques.
Congenital heart disease (CHD) is the most common To address the limitations of both manual processing
cause of infant mortality, affecting 2.4 to 13.7 per 1000 and existing deep learning models, we propose a semi-
newborns. Obstetric ultrasound (US) has been the automatic segmentation pipeline that introduces two key
1
gold standard imaging method for the detection and innovations. Firstly, a region-limited Faster Region-based
diagnosis of fetal malformations during the antenatal Convolutional Neural Network (R-CNN)-based position
period. Initially, the International Society of Ultrasound in detector is applied to isolate fetal heart structures from
Obstetrics and Gynecology (ISUOG) emphasized the use noisy US backgrounds, reducing false positives caused
of four-chamber view (FCV) in fetal cardiac examinations. by rib shadowing and organ overlap. Secondly, instead of
Subsequent guidelines introduced the outflow tract view
(OTV) and three-vessel view (TVV) to visualize the overall relying solely on fully automated thresholding, the system
cardiac structure for a comprehensive assessment. Recent integrates expert-in-the-loop adjustments, allowing
2
advancements in 3D US techniques have significantly sonographers to interactively fine-tune segmentation based
enhanced the visualization of spatial structures of the fetal on grayscale analysis. This hybrid strategy offers improved
heart. However, 3D structures cannot be comprehended interpretability, flexibility, and clinical usability—filling a
3
thoroughly based on 2D images, which usually leads to critical gap between automated artificial intelligence (AI)
low-efficiency treatment strategy planning for the fetus. 4 tools and practical medical workflow.
In medical imaging, deep learning approaches have In this study, we introduced the fetal heart reconstruction
5
6
gained prominence in recent years. Mofrad et al. utilized technique (FRT) for the semi-automatic segmentation
deep learning to augment and assist clinicians in the and post-processing of fetal echocardiography (Figure 1).
automated diagnosis of adult heart diseases. Nonetheless, Initially, we converted US volume scans into sequential
constructing accurate fetal heart models from US data 2D images and performed detailed frame-by-frame
remains challenging due to the small heart size. To address segmentation of the fetal heart using FCV, OTV, and TVV
this, we employed deep learning for precise fetal heart to train and test the R-CNN. The R-CNN then identified
modeling, aiding in diagnosis and prenatal screening. the fetal heart’s position within the echocardiogram
Moreover, 3D printing technology offers opportunities to data and generated a segmentation mask. The FRT was
create tangible fetal heart models, benefiting various facets of subsequently employed to perform segmentation of the
fetal healthcare such as surgeon training, personalized pre- US images, generating results for reconstructing a basic
surgical planning, and doctor–patient communication. 7–10 digital model of the fetal heart. Ultimately, these digital
In this study, we employed PolyJet multi-material fetal heart models were 3D printed. This method offers
3D printing due to its high resolution (~200 µm), a semi-automated approach to rapidly reconstruct fetal
excellent surface finish, and ability to reproduce small heart models from US data, presenting a new diagnostic
and complex anatomical structures such as fetal heart strategy for clinical fetal heart analysis.
chambers and vessels. Compared with fused deposition
modeling (FDM) or stereolithography (SLA), PolyJet 2. Methods
offers greater accuracy and material flexibility, making 2.1. Study participants
it particularly well-suited for fabricating detailed fetal The study was approved by the Ethics Committee of the
cardiac models for clinical and educational applications. Second Affiliated Hospital of Air Force Medical University
In parallel, the integration of machine learning into (approval no. TDLL-202402-01). All pregnant women in
additive manufacturing processes has garnered increasing this study were informed of the safety and limitations of US
11
attention. Zhang et al. provided a comprehensive review and signed an informed consent form prior to examination.
of how neural networks and reinforcement learning are We retrospectively studied the echocardiographic volume
being utilized to optimize printing parameters, material datasets obtained by 4D US with Spatiotemporal Image
formulations, and fabrication efficiency. Similarly, Li et al. Correlation from 100 normal fetuses as compulsory
12
highlighted the role of big data and digital twin frameworks supervised learning cases. All of the fetuses were singletons
in enabling adaptive and intelligent control of 3D printing and were selected from prenatal screening tests in the
workflows. While these efforts are primarily focused Department of Ultrasound Diagnosis at Tangdu Hospital
on manufacturing engineering, our work addresses the (China) from January 2019 to May 2021.
upstream phase—employing deep learning to enhance
the fidelity of anatomical models derived from US data, 2.2. Data acquisition
which ultimately contributes to higher-quality medical 3D The scanning operation was strictly performed in
printing outputs. However, the time and resources required accordance with the practice guidelines updated by the

Volume 11 Issue 4 (2025) 243 doi: 10.36922/IJB025200192