USFetal: Tools for Fetal Brain Ultrasound Compounding

TL;DR

USFetal introduces a systematic taxonomy of fetal brain ultrasound compounding methods across four families—multi-scale, transformation-based, variational, and deep learning—emphasizing unsupervised and self-supervised strategies due to the lack of artifact-free ground truth. It implements representative methods from each class, including a DL-based self-supervised framework and a training-free plug-and-play priors approach, and benchmarks them on ten multi-view fetal ultrasound datasets with expert radiologist scores and standard metrics. The study reveals that conventional quantitative metrics do not reliably reflect perceptual or diagnostic quality in fetal ultrasound, underscoring the value of expert assessment and task-aware evaluation. The work provides a publicly available USFetal Compounding Toolbox to enable benchmarking and future research, and suggests future directions toward unified registration-compounding frameworks and integration with higher-level clinical tasks.

Abstract

Ultrasound offers a safe, cost-effective, and widely accessible technology for fetal brain imaging, making it especially suitable for routine clinical use. However, it suffers from view-dependent artifacts, operator variability, and a limited field of view, which make interpretation and quantitative evaluation challenging. Ultrasound compounding aims to overcome these limitations by integrating complementary information from multiple 3D acquisitions into a single, coherent volumetric representation. This work provides four main contributions: (1) We present the first systematic categorization of computational strategies for fetal brain ultrasound compounding, including both classical techniques and modern learning-based frameworks. (2) We implement and compare representative methods across four key categories - multi-scale, transformation-based, variational, and deep learning approaches - emphasizing their core principles and practical advantages. (3) Motivated by the lack of full-view, artifact-free ground truth required for supervised learning, we focus on unsupervised and self-supervised strategies and introduce two new deep learning based approaches: a self-supervised compounding framework and an adaptation of unsupervised deep plug-and-play priors for compounding. (4) We conduct a comprehensive evaluation on ten multi-view fetal brain ultrasound datasets, using both expert radiologist scoring and standard quantitative image-quality metrics. We also release the USFetal Compounding Toolbox, publicly available to support benchmarking and future research. Keywords: Ultrasound compounding, fetal brain, deep learning, self-supervised, unsupervised.

Figures (4)

• Figure 1: Schematic illustration of multi-view fetal brain ultrasound acquisition and compounding. Each acquired 3D volume, ${y}_i$, captures a partial view of the brain, where the varying probe orientation determines the viewing angle and view-dependent artifacts. The views are typically not aligned and differ significantly in appearance. Compounding aims to integrate these complementary, imperfect views into a coherent, high-quality representation of the whole brain.
• Figure 2: Schematic of the proposed self-supervised compounding architecture. The input volumes $\{y_i\}_{i=1}^N$ are first processed by an edge extractor module, which captures anatomical boundaries and forwards them via skip connections. Each view is then projected into a shared feature space and encoded across multiple spatial scales. The encoded features are fused and decoded to form a preliminary reconstruction, which is subsequently refined by integrating the skipped edge information. The resulting compounded volume is compared to the input views using a self-supervised loss that enforces both voxel-wise and structural consistency, enabling end-to-end optimization without the need for ground-truth supervision.
• Figure 3: Compounding results from five representative methods applied to a fetal brain ultrasound dataset acquired from a single subject (Subject 8, eight 3D input views). The upper panel shows the input multi-view ultrasound volumes, and the lower panel displays the compounded outputs generated by the DoG, PCA, Variational, PnP, and SSL methods. Qualitative expert ratings (1–3 scale; higher indicates better perceptual quality) for this subject were: DoG = 3, PCA = 1, Variational = 2, PnP = 1, and SSL = 2.
• Figure 4: Compounding results from five representative methods applied to a fetal brain ultrasound dataset acquired from a single subject (Subject 5, five 3D input views). The upper panel shows the input multi-view ultrasound volumes, and the lower panel displays the compounded outputs generated by the DoG, PCA, Variational, PnP, and SSL methods. Qualitative expert ratings (1–3 scale; higher indicates better perceptual quality) for this subject were: DoG = 3, PCA = 2, Variational = 1, PnP = 2, and SSL = 2.