L-FUSION: Laplacian Fetal Ultrasound Segmentation & Uncertainty Estimation

Johanna P. Müller, Robert Wright, Thomas G. Day, Lorenzo Venturini, Samuel F. Budd, Hadrien Reynaud, Joseph V. Hajnal, Reza Razavi, Bernhard Kainz · 2025

Read L-FUSION: Laplacian Fetal Ultrasound Segmentation & Uncertainty Estimation by Johanna P. Müller, Robert Wright, Thomas G. Day, Lorenzo Venturini, Samuel F. Budd, Hadrien Reynaud, Joseph V. Hajnal, Reza Razavi, Bernhard Kainz (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-05245.txt.

Introduction

Johanna Müller and co-authors present, in twenty twenty-five, L-FUSION, a tool to mark out fetal ultrasound scans and gauge doubt in the result. Ultrasound helps find growth problems early, before birth. But scan quality shifts with the skill of the person who does the scan. Flaws in the image also make scans hard to read. L-FUSION joins large foundation models with ways to gauge doubt. It uses spread from stochastic segmentation networks and fast Laplace approximations, a math shortcut. That lets it gauge one type of doubt from the marked output alone. A dropout part is built in too.

This setup tells lesions apart from normal anatomy with no hand labels for disease. It gives better doubt maps and what-if views for instant feedback. Tests across data sets showed top marks for how well it marked scans and steady doubt results.

Catalog record

CREATORJohanna P. Müller, Robert Wright, Thomas G. Day, Lorenzo Venturini, Samuel F. Budd, Hadrien Reynaud, Joseph V. Hajnal, Reza Razavi, Bernhard Kainz
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-05245

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

L-FUSION: Laplacian Fetal Ultrasound Segmentation & Uncertainty Estimation Johanna P. Müller, Robert Wright, Thomas G. Day, Lorenzo Venturini, Samuel F. Budd, Hadrien Reynaud, Joseph V. Hajnal, Reza Razavi, Bernhard Kainz Abstract Accurate analysis of prenatal ultrasound (US) is essential for early detection of developmental anomalies. However, operator dependency and technical limitations ( e.g. intrinsic artefacts and effects, setting errors) can complicate image interpretation and the assessment of diagnostic uncertainty. We present L-FUSION (Laplacian Fetal US Segmentation with Integrated FoundatiON models), a framework that integrates uncertainty quantification through unsupervised, normative learning and large-scale foundation models for robust segmentation of fetal structures in normal and pathological scans. We propose to utilise the aleatoric logit distributions of Stochastic Segmentation Networks and Laplace approximations with fast Hessian estimations to estimate epistemic uncertainty only from the segmentation head. This enables us to achieve reliable abnormality quantification for instant diagnostic feedback. Combined with an integrated Dropout component, L-FUSION enables reliable differentiation of lesions from normal fetal anatomy with enhanced uncertainty maps and segmentation counterfactuals in US imaging. It improves epistemic and aleatoric uncertainty interpretation and removes the need for manual disease-labelling. Evaluations across multiple datasets show that L-FUSION achieves superior segmentation accuracy and consistent uncertainty quantification, supporting on-site decision-making and offering a scalable solution for advancing fetal ultrasound analysis in clinical settings. Code is available at https://github.com/ividja/L-FUSION . 1 Introduction Fetal ultrasound (US) is a key tool for monitoring development and detecting abnormalities early, with detection rates improving from 15\text{--}59\% in the 2000s to 30\text{--}88\% today, depending on the lesion type [ 6 ] . However, US imaging remains challenging due to noise, shadows, and anatomical ambiguity. Machine learning models can support clinicians by providing consistent and objective assessments, with detection performance often matching expert evaluation [ 5 , 24 ] . Yet, standard deterministic models offer no confidence estimates, limiting trust and interpretability in safety-critical settings. Uncertainty quantification helps flag unreliable predictions and guide clinical attention. Current approaches, however, face two main challenges: (1) high computational cost, especially for high-dimensional imaging [ 29 ] , and (2) the need for large, diverse datasets to capture epistemic uncertainty, critical for identifying unfamiliar or out-of-distribution (OOD) cases [ 12 , 15 ] . Laplacian approximations offer a computationally efficient method for capturing both epistemic and aleatoric uncertainty, especially when paired with fast Hessian estimates. Meanwhile, pretrained US foundation models [ 13 ] address data scarcity by enabling transfer learning across domains. When integrated, these components provide accurate, generalisable segmentation with reliable uncertainty estimates, supporting robust and scalable diagnostic systems for real-world fetal US. Contributions. We propose L-FUSION , a novel framework for fetal ultrasound segmentation that integrates foundation model embeddings with two complementary uncertainty-aware heads. Our key contributions are: (1) We combine a deterministic ultrasound foundation encoder with a Laplacian head for calibrated uncertainty and a Dropout head for counterfactual segmentations, enabling joint aleatoric and epistemic uncertainty quantification. This improves segmentation accuracy and unsupervised OOD detection. (2) We reduce computational cost by applying Laplace approximation only to the segmentation head and using efficient Hessian approximations. The frozen encoder and lightweight heads ensure scalability across diverse clinical ultrasound tasks. (3) We introduce the first 3D Laplacian segmentation model for ultrasound clips, enabling spatiotemporal extension of our method. In the absence of 3D foundation models, we leverage the 2D encoder as a proxy. Related Work. Automated ultrasound segmentation lags behind other imaging modalities due to inconsistent metrics, smaller datasets, and high inter- and intra-expert variability [ 20 ] . However, growing emphasis on rigorous validation, cross-method comparisons, and larger standardised datasets aims to support clinical uptake [ 20 ] . Recent advances incorporate systematic priors, and geometric, temporal, and physics-based constraints to boost accuracy [ 19 , 26 , 16 , 9 ] . Foundation models pretrained on large-scale natural images have been adapted for medical imaging, including ultrasound, though cross-modality generalisation remains limited [ 28 ] . The Multi-Organ FOundation (MOFO) model [ 2 ] enhances segmentation by learning organ-invariant representations and anatomical priors, outperforming single-organ approaches. Similarly, the Universal US Foundation Model (USFM) [ 13 ] improves generalisation across organs, devices, and centres via self-supervised spatial-frequency dual-masked image modelling on a large, diverse dataset. While foundation models offer rich features, they lack inherent uncertainty quantification. Bayesian methods provide probabilistic segmentation, estimating epistemic and aleatoric uncertainty [ 1 , 3 , 8 ] . Laplace approximations efficiently estimate epistemic uncertainty by approximating the posterior, avoiding costly Hessian computation [ 4 ] . Recent work combines aleatoric logit distributions from Stochastic Segmentation Networks [ 18 ] with Laplace approximations for comprehensive uncertainty quantification [ 27 ] . Figure 1 : Foundation model (Encoder) with Laplacian segmentation head L(f) and a Dropout segmentation path. Here, L(f) and Dropout Segmentation Head G(f) decoding the embeddings f=\bar{F}(x) . The segmentations S_{i}(f) are predicted by L(f) . We enhance the uncertainty measure maps U_{L/D,k} with Uncertainty-Calibration as U_{fused} by including the variance and total entropy of logits. 2 Method We introduce L-FUSION, a probabilistic segmentation framework for fetal ultrasound that combines a frozen foundation model, e.g. , USFM [ 13 ] , with two complementary uncertainty-aware segmentation heads: a Laplacian-based branch and a Dropout-based branch [ 7 ] . By jointly modelling both epistemic and aleatoric uncertainty, L-FUSION produces more reliable predictions and highlights regions of ambiguity, particularly useful for clinical interpretation and decision-making. An overview of the framework is shown in Fig. 1 . Feature Encoding L-FUSION begins by passing an input 2D ultrasound image x through a pretrained foundation encoder \bar{F} , e.g. , USFM [ 13 ] . The encoder remains frozen during training to preserve the generalisable representations learned during pretraining and to reduce the computational burden associated with fine-tuning the encoder as well. The output feature embedding f=\bar{F}(x) is shared by both segmentation branches. This design choice decouples representation learning from downstream uncertainty estimation, ensuring that all uncertainty arises from the segmentation heads rather than encoder variability. Theorem 2.1 (Sufficiency of Pretrained Embeddings for Segmentation and Uncertainty Quantification) Let x\in\mathcal{X} be an input image, and let \bar{F}:\mathcal{X}\to\mathcal{F} be a deterministic, pretrained foundation model mapping x to a feature embedding f=\bar{F}(x) . Consider a Laplacian segmentation head L that models the posterior segmentation distribution as p(S|f,\theta)=\prod_{i}\text{Bernoulli}(s_{i}|\sigma(\eta_{L,i})), where the latent logits \eta follow a Laplace-approximated posterior: p(\eta|f,\theta)=\mathcal{N}(\eta|\mu_{\theta}(f),PP^{\top}+D), with \mu_{\theta}(f) the mean prediction, and PP^{\top}+D the covariance matrix. If \bar{F} preserves all segmentation-relevant information, i.e., using the mutual information I(S;x)=I(S;f), then f is a sufficient statistic for estimating segmentation and associated predictive uncertainty. Proof Since \bar{F} is deterministic, the distribution over f given x reduces to a delta function: p(f|x)=\delta(f-\bar{F}(x)), so the segmentation posterior satisfies p(S|x)=p(S|f,\theta). By the Data Processing Inequality [ 23 ] , I(S;x)\geq I(S;f). We assume I(S;x)=I(S;f), i.e., that no information about S is lost in passing from x to f . This implies that f is a sufficient statistic for S with respect to x . Now, consider uncertainty estimation. The Laplace approximation assumes a Gaussian posterior over parameters: q(\theta^{*})=\mathcal{N}(\theta^{*}|\theta_{\text{MAP}},\mathbf{H}^{*-1}), where \mathbf{H}^{*} is the Hessian of the negative log-likelihood at the MAP estimate. Predictive uncertainty is then propagated via linearisation of the model output \eta around \theta_{MAP} , yielding: \operatorname{Var}(S|f)\approx J\mathbf{H}^{*-1}J^{\top}+\sigma^{2}(f)I, where J is the Jacobian of

Read the complete text