Developing and Testing AI Models for Fetal Biometry and Amniotic Volume Assessment in Fetal Ultrasound Scans.

Overview

Routine fetal ultrasound scan during the second trimester of the pregnancy is a low-cost, noninvasive screening modality that has been proven to lower fetal mortality by up to 20%. One of the critical elements of this exam is the measurement of fetal biometric parameters, which are the head circumference (HC), biparietal diameter (BPD), abdominal circumference (AC), and femur length (FL) measured on biometry standard planes. Those standard planes are taken according to quality standards first described by Salomon et al. and used as the guidelines of the International Society of Ultrasound in Obstetrics and Gynecology (ISUOG). The biometric parameters extracted from them are essential to diagnose fetal growth restriction (FGR), the world's first cause of perinatal fetal mortality. Such measurements and image quality assessment are time-consuming tasks that are prone to inter and intraobserver variability depending on the level of skill of the sonographer or the physician performing the exam. Amniotic fluid (AF) volume assessment is also an essential step in routine screening scans allowing the diagnosis of oligo or hydramnios, both associated with increased fetal mortality rates. The AF is measured by two main "semi-quantitative" techniques: Amniotic Fluid Index (AFI) and the single deepest pocket (SDP). The latter is more specific as it lowers the overdiagnosis of oligo-amnios without any impact on mortality or morbidity and is easier to perform for the sonographer (only one measurement versus four in the case of the AFI technique). However, AF assessment remains a time-consuming and poorly reproducible task. Attempts to automate such biometric measurements and AF volume assessment have been made using Artificial Intelligence (AI) and deep learning (DL) tools. Studies showed excellent results "in silico," reaching up to 98 %, 95%, 93 % dice score coefficients for HC, AC, and FL measurements and 89 % DSC for AFI measurements. However, they were all conducted retrospectively without validation on prospectively acquired images. Reviews and experts have stressed the need for quality peer-reviewed prospective studies to assess AI tools' performance with real-world data. Their performance is expected to be worse and to reflect better their use in the clinical workflow. This study aims to develop DL models to automate HC, BPD, AC, and FL measurements and AF volume assessment from retrospectively acquired data and test their performances to those of clinicians and experts on prospective real-world fetal US scans.

The DL models will be trained, validated, and tested on the retrospectively acquired data first. This data will consist of fetal US images gathered in the participating medical centers after patient-level anonymization. The ground truth for the models will consist of annotations made by radiologists and obstetricians for classification and segmentation purposes. The DL models will be trained to perform the following tasks: * Detection of the following standard planes as described in the ISUOG guidelines: transthalamic, transventricular, transcerebellar, abdominal, and femoral planes on video loops. * Image quality scoring according to the ISUOG guidelines of the transthalamic, abdominal and femoral planes. * Fetal cranium, abdomen, and femur segmentation to measure HC, BPD AC, and FL. * Detection of AF pockets. * Segmentation of AF pockets and extraction of pockets depth in order to evaluate the SDP measurement Physicians will be asked to save additional images and video loops additional to their routine screening in the prospective examinations: * Eight images: transthalamic, abdominal, and femoral standard planes with and without calipers, SDP with and without calipers. * Four video loops up to five seconds each: * A cephalic loop encompassing the transcerebellar, transthalamic, and transventricular planes. * An abdominal loop going from the four-chamber view of the heart to a cross-section of the kidneys and back. * A femoral loop with the probe parallel to the sagittal axis of the femur sweeping from side to side. * A whole amniotic cavity loop, with the probe perpendicular to the ground applying as little pressure as possible on the patient's abdomen, sweeping from the uterine fundus to the cervix, once or twice depending on the volume of the amniotic cavity. The clinicians performing the exam in "real-time"(RT clinicians), the panel of experts, and the DL models will review the prospective examinations. The SDP measurement extracted by the AF pocket detection and segmentation models will be directly compared to the value measured by the RT clinicians. Then, the image quality of planes selected by the RT clinicians and the model will be scored by the panel of experts. The segmentation task will be evaluated in a tripartite fashion: the model, the RT clinicians, and the panel will all segment the same images. To assess inter-observer agreement, 10% of the images will be randomly selected and reviewed by two independent reviewers from the panel.