3DFETUS: Deep Learning-Based Standardization of Facial Planes in 3D Ultrasound

TL;DR

This work tackles the variability in 3D fetal facial plane analysis by introducing GT++ to construct robust, landmark-based ground-truth planes and 3DFETUS, a differentiable DL framework that standardizes fetal facial orientation to a canonical pose. GT++ improves landmark completion and plane estimation, reducing inter-observer variability and enhancing clinical acceptability. 3DFETUS leverages a three-branch feature extractor, 6D pose regression, and a spatial transformer to align volumes with a grid-based loss, achieving superior rotation and translation accuracy versus state-of-the-art methods while maintaining computational efficiency. The combined pipeline demonstrates strong qualitative and quantitative performance across gestational ages, reinforcing its potential for standardized, low-cost fetal facial evaluation in routine clinical practice.

Abstract

The automatic localization and standardization of anatomical planes in 3D medical imaging remains a challenging problem due to variability in object pose, appearance, and image quality. In 3D ultrasound, these challenges are exacerbated by speckle noise and limited contrast, particularly in fetal imaging. To address these challenges in the context of facial assessment, we present: 1) GT++, a robust algorithm that estimates standard facial planes from 3D US volumes using annotated anatomical landmarks; and 2) 3DFETUS, a deep learning model that automates and standardizes their localization in 3D fetal US volumes. We evaluated our methods both qualitatively, through expert clinical review, and quantitatively. The proposed approach achieved a mean translation error of 3.21 $\pm$ 1.98mm and a mean rotation error of 5.31 $\pm$ 3.945$^\circ$ per plane, outperforming other state-of-the-art methods on 3D US volumes. Clinical assessments further confirmed the effectiveness of both GT++ and 3DFETUS, demonstrating statistically significant improvements in plane estimation accuracy.

Figures (9)

• Figure 1: Overview of the proposed 3DFETUS framework for automatic standardization of fetal facial planes in 3D US. The pipeline consists of two main stages: (1) GT++ Ground Truth Construction (top row), where the facial ground truth planes are derived from anatomical landmarks on 3D US volumes, and (2) 3DFETUS Network Architecture (bottom row), a deep learning model comprising three blocks: a feature extractor, affine transformation regressor (rotation and translation), and a spatial transformer module. The network processes three orthogonal US slices as input and predicts the transformation needed to align the fetal face into a canonical frontal pose. The spatial transformer then warps the original 3D volume accordingly, producing a standardized US volume and aligned sagittal, coronal, and axial facial planes.
• Figure 2: Illustration of the anatomical landmarks, canonical facial pose and facial planes considered. Landmark abbreviations: Landmark abbreviations: exR, exL = exocanthion right, left; enR, enL = endocanthion right, left; n = nasion; aR, aL = alare right, left; acR, acL = alar crest right, left; prn = pronasale; sn = subnasale; chR, chL = cheilion right, left; cphR, cphL = crista philtrum right, left; ls = labiale superius; li = labiale inferius; sl = sublabiale; pg = pogonion; tR, tL = tragion right, left; oiR, oiL = otobasion inferius right, left.
• Figure 3: Examples of facial planes generated using GT-base and GT++. Clinician-defined landmarks are shown in purple, while red landmarks represent those added through the completion process. Each row corresponds to a different subject. The columns display sagittal, coronal, and axial views, for both methodologies to facilitate comparison.
• Figure 4: Boxplots of rotation (a) and translation (b) errors on the test set. Comparison of 3DFETUS, ITN Li2018 , and acquisition error relative to GT facial planes.
• Figure 5: 3DFETUS performance on the test set. Displayed are 2D planes from three different test subjects representing high, average, and low performance of the proposed 3DFEUS method. For each subject, the Euclidean angle (EA, in degrees) and translation error ($t$, in mm) are reported with respect to the ground truth (GT). The mean quality score ($z$), as assigned by clinicians, is also indicated.