Anatomy-Guided Foundation Model Adaptation with Within-Case Prototype Supervision for Standard Plane Detection in Fetal Ultrasound Blind Sweeps

Yuzhe Zhao

cs.CVAug 27, 2026

Anatomy-Guided Foundation Model Adaptation with Within-Case Prototype Supervision for Standard Plane Detection in Fetal Ultrasound Blind Sweeps

Authors: Yuzhe Zhao

Organizations: School of Computing, University of Leeds, Leeds, UK

Abstract

Detecting the fetal abdominal circumference standard plane in low-cost obstetric blind sweeps is a highly imbalanced frame-classification problem: positive frames account for under 3% of a sequence, form short contiguous segments, and are poorly handled by off-the-shelf ultrasound and vision foundation models. We propose AnatoProto, a lightweight sequence-level framework that adapts a frozen BiomedCLIP encoder to fetal blind sweeps through four components: (i) anatomy-weighted spatial pooling that uses nnU-Net abdominal-region probabilities as a spatial prior to reweight BiomedCLIP patch tokens, so frozen semantic features are aggregated onto anatomically meaningful regions; (ii) a within-case prototype loss that pulls each frame embedding toward the mean of positive frames of the same sweep, exploiting case-level structure unavailable at the frame level; (iii) a three-stage cascade refinement (frame->segment->case-level rejecter) that lifts the prediction unit from noisy frames to structurally-constrained segments; and (iv) a hybrid prediction head that jointly models per-frame stability and inter-frame boundary transitions to suppress boundary false positives. On the ACOUSLIC-AI benchmark, AnatoProto reaches a test F1 of 67.72, outperforming the strongest foundation-model baseline (FetalCLIP + PRS, F1 = 54.52) by +13.20 F1 and the strongest video temporal-action-detection baseline (TriDet + PRS) by +15.76 F1. A synergy study, backed by embedding geometry and paired-bootstrap confidence intervals, shows that the prototype loss and anatomy-weighted pooling are not additive: applied alone the prototype loss reduces recall by 12 points, but combined with anatomy-weighted pooling it increases recall by 6.5 points -- a sign-flip we trace to the accuracy of the within-case prototype.

Figures & tables

Figure 1: Overview of AnatoProto . Top (training, per sweep): each frame of a blind sweep (real example frames shown) is encoded by a frozen BiomedCLIP ViT-B/16 into $7\times 7$ patch tokens, and by a frozen single-organ nnU-Net into a $7\times 7$ abdominal probability map that serves as a spatial prior reweighting the semantic tokens (anatomy-weighted pooling; zero new parameters). A 4-layer bi-directional temporal transformer (BiTT) consumes the pooled sequence and drives a frame head $\hat{p}_{t}$ , a stability head $s_{t}$ and a boundary head $b_{t}$ ; a within-case prototype loss pulls the positive embeddings of the same sweep onto a shared attractor and pushes negatives away. Bottom (inference, coarse-to-fine): per-frame probabilities of a real test sweep are binarised by PRS into structurally-constrained segments, and a logistic-regression rejecter discards implausible ones; the final segments (green) closely match the ground truth (blue).

Feature	$\bar{s}$	seg. len.	$\max s$	$\mathrm{ratio}_{s>0.5}$	$\bar{b}$	$\max b$	centre	$\mathrm{std}\,s$	(intercept)
Weight	$+1.76$	$+1.30$	$+0.76$	$-0.55$	$-0.41$	$+0.19$	$-0.12$	$+0.06$	$+0.58$

Table 1: Standardised logistic-regression coefficients of the Stage-3 segment rejecter (class-balanced,

\lambda_{\ell_{2}}=1

), sorted by

|w|

. Positive weight

\Rightarrow

“keep”; negative

\Rightarrow

“reject”.

Group	Method	Precision	Recall	F1
Official	ACOUSLIC nnU-Net [ 11 ]	10.41	20.39	13.78
Frame-based CNN	ResNet-50 [ 6 ]	–	–	9.80
Frame FM	USFM [ 16 ] + linear probe + PRS	5.74	35.96	9.90
Frame FM	DINOv3-L/16 [ 9 ] + linear probe + PRS	7.29	16.60	10.13
Frame FM	SonoCLIP [ 14 ] + cls head + PRS	6.22	15.78	8.92
Frame FM	FetalCLIP-CLS [ 8 ] + LoRA + PRS	21.52	19.19	20.29

Table 2: Comparison with state-of-the-art on our

45

-case test partition of ACOUSLIC-AI. All numbers are frame-level and averaged over cases. Best in bold . “PRS” denotes the temporal post-processing of Sec. 3.4 Stage 2; it is applied to each external frame FM at the threshold that maximises its own F1 on the validation split, and reported without PRS where it did not improve (e.g., BiomedCLIP + linear probe, ResNet-50).

Figure 2: Frame-level precision–recall curves on the

45

test cases (points pooled across sweeps). Our D-3 configuration and its boundary-head variant sit an order of magnitude above the strongest external single-frame FM baseline (FetalCLIP + linear probe) in average precision, and the ordering holds across the entire recall range.

Configuration	Precision	Recall	F1
Baseline (BiomedCLIP + uniform pool)	49.5	71.1	58.37
+ Prototype loss (only)	57.6 ( $\checkmark$ +8.1)	59.0 ( $\checkmark$ $-$ 12.0)	58.30
+ Anatomy-weighted pool (only)	60.0 ( $\checkmark$ +10.6)	65.4 ( $-$ 5.7)	62.58
+ Both = D-3	59.3 ( $\checkmark$ +9.8)	71.9 (+0.8)	65.03

Table 3: Ablation of the two components of the D-3 configuration. All bracketed differences are computed vs. the baseline row ; a

\checkmark

marks those whose

95\%

paired-bootstrap CI does not cross zero. The critical sign flip (Recall

-12.0

on baseline vs.

+6.5

on anatomy-alone) is a marginal comparison and is discussed in the text.

Configuration	margin_sat +	margin_sat -	between
Pre-training (frozen BiomedCLIP, pooling only)
Uniform pool	0.537	–	0.257
Anatomy-weighted pool	0.645	–	0.308
Post-training (BiTT + hybrid loss)
Baseline (uniform pool, BCE only)	0.252	0.282	0.144
+ Prototype loss (only)	0.666	0.722	0.433

Table 4: Embedding-geometry evidence at two stages of training. margin_sat ± = fraction of positive / negative frames satisfying the corresponding margin; between = between-class prototype distance. The pre-training panel is computed on the frozen BiomedCLIP embeddings without any temporal-head training; margin_sat - is undefined pre-training because it requires a trained negative attractor.

Configuration	F1	$\Delta$ vs. D-3
D-3	65.03	–
+ Boundary head	65.61	+0.58
+ Stability head	65.87	+0.84
+ Cascade rejecter	66.12	+1.09
+ Cascade + Stability	66.38	+1.35
+ Cascade + Boundary	66.52	+1.49

Table 5: Cascade / hybrid-head ablation, all measured on top of D-3.

Figure 3: Qualitative comparison of predicted positive segments on four representative test sweeps ( $x$ -axis: frame index $0$ – $840$ ). Blue : ground truth. Red : FetalCLIP + PRS (strongest external baseline). Green : AnatoProto (full model). (a) Best-gain sweep – our method perfectly localises the abdominal-circumference segment while FetalCLIP scatters detections. (b) Typical sweep with multiple ground-truth segments – our method recovers all three while FetalCLIP recovers only one. (c) Easier sweep – comparable performance. (d) Hard sweep – both methods fail, but our method makes no prediction while FetalCLIP produces two false-positive segments.

Figure 4: Per-frame probability traces on two multi-segment test sweeps. Light-blue bands = ground-truth positive segments; red blocks = uniform-pool baseline binary predictions; green curve = AnatoProto (D-3) per-frame probability; dark-green blocks = AnatoProto binary predictions at

\tau=0.55

. The baseline scatters detections far from any ground-truth plane in (A) and misses the third segment while firing a spurious one in (B); AnatoProto produces localised probability peaks that align with all ground-truth segments – exactly the failure modes the anatomy-weighted pool is designed to close. Per-case metrics are printed above each panel.

Figure 5: Visual detection results on two test sweeps (A: e3670fa9 ; B: ba245b67 ). Columns: ground-truth keyframe; top-scoring frame of the uniform-pool baseline; top-scoring frame of AnatoProto (D-3). The green contour is the annotated fetal abdomen and is present only when the selected frame is a true AC plane. The baseline’s top frame is a temporally misaligned look-alike without abdominal anatomy (frames 731 / 517), whereas AnatoProto lands inside the ground-truth segment (frames 319 / 204). Timeline insets: per-sweep ground truth (blue), baseline predictions (red), our predictions (green); triangles mark the top-scoring frames.

Ultrasound Imaging

Ultrasound imaging uses sound waves to visualize anatomy and function, with AI improving image reconstruction, diagnosis, measurement, and scan guidance.

See all 150 papers

May 27, 2026

Benchmarking Ultrasound Foundation Models for Fetal Plane Classification

Ultrasound is widely used in obstetric care due to its safety, accessibility, and real-time imaging. However, interpretation remains operator-dependent and susceptible to noise and artifacts. Deep learning models have shown strong performance to solve these problem, but they typically require large annotated datasets that are difficult to obtain in clinical ultrasound. Foundation models (FMs) offer an alternative, using a large number of ultrasound images to learn transferable representations that can generalize with limited labeled data. This work presents a comprehensive benchmark of ultrasound-specific FMs for fetal plane classification. We evaluated four ultrasound FMs (USFM, MOFO, UltraSAM, FetalCLIP) against two CNN baselines (ResNet50, EfficientNet-V2) and a ViT (DINOv3) pretrained on natural images. We trained all models under two complementary settings: full fine-tuning and linear probing with a frozen encoder. All models were trained using 5-fold patient-level cross-validation on a Spanish fetal ultrasound dataset and tested on both in-domain data and an external African cohort to assess cross-population generalization. We found that FetalCLIP achieved the best results in the linear probing setting (F1 = 0.9261 for in-domain, F1 = 0.9731 for out-of-domain), while USFM performed best in the full fine-tuning setting (F1 = 0.9476 for in-domain, F1 = 0.9515 for out-of-domain). MOFO and UltraSAM degraded most in both settings, underperforming natural image pretrained models in some cases. These findings highlight how the choice of pretrained model strongly affects fetal plane classification performance, since different pretraining objectives lead to different levels of transferability.

Leya Barrientos, Yuexi Du, Nicha C. Dvornek

Radiology & Biomedical Imaging, Yale School of Medicine, USA · Department of Biomedical Engineering, Yale University, USA

Medical Imaging Foundation Models Medical Image Classification Ultrasound Imaging Medical Image Benchmarks Ultrasound Image Classification

Ask PDF

Apr 22, 2026

Structure-Augmented Standard Plane Detection with Temporal Aggregation in Blind-Sweep Fetal Ultrasound

In low-resource settings, blind-sweep ultrasound provides a practical and accessible method for identifying fetal growth restriction. However, unlike freehand ultrasound which is subjectively controlled, detection of biometry plane in blind-sweep ultrasound is more challenging due to the uncontrolled fetal structure to be observed and the variaties of oblique planes in the scan. In this work, we propose a structure-augmented system to detect fetal abdomen plane, where the abdominal structure is highlighted using a segmentation prior. Since standard planes are emerging gradually, the decision boundary of the keyframes is unstable to predict. We thus aggregated the structure-augmented planes with a temporal sliding window to help stabilise keyframe localisation. Extensive results indicate that the structure-augmented temporal sliding strategy significantly improves and stabilises the detection of anatomically meaningful planes, which enables more reliable biometric measurements in blind-sweep ultrasound.

Keli Niu, He Zhao, Qianhui Men

University of Bristol · University of Liverpool

Ultrasound Imaging

Ask PDF

Jul 1, 2026

Foundation Model-driven Key Anatomy Frame Selection for Blind-sweep Ultrasound Fetal Birth Weight Estimation

Accurate fetal birth weight (FBW) estimation shortly before delivery is clinically valuable yet challenging due to its reliance on operator expertise, particularly in low-resource settings. To reduce this reliance, we study near-term birth-weight regression from blind-sweep ultrasound (US) videos acquired within 48 hours prior to delivery, with post-delivery weighing as ground truth. Accordingly, we propose a foundation model-driven key anatomy frame selection framework that enables accurate FBW regression despite the absence of plane constraints in blind sweeps. Our highlights are as follows: (1) We believe this is the first work to estimate FBW using blind-sweep US videos, enabling operator-independent assessment. (2) An Anatomy-Guided Frame Selection module equipped with a vision-language foundation model is proposed for keyframe collection in unconstrained sweeps. (3) A Redundancy-Aware Feature Compression module is designed to compress frame features while preserving task-relevant information, alleviating temporal redundancy. Extensively validated on prospectively collected data from 839 patients, our method achieves an MAE of 161.3 g, with 90.23% and 100% of cases falling within 10% and 15% absolute percentage error, outperforming typical Hadlock estimation and strong competitors. Codes are available at https://github.com/ouleoule/BlindSweep-EBW.

Le Ou, Xiliang Zhu, Huanwen Liang +8

Medical Ultrasound Image Computing (MUSIC) Lab, Shenzhen University, Shenzhen, China · School of Biomedical Engineering, Shenzhen University, Shenzhen, China · School of Biomedical EngineeringNanjing,and Informatics,China Nanjing Medical University +4

Ultrasound Imaging Video Frame Selection

Ask PDF

Abstract

Ultrasound Imaging

Explore similar work

Benchmarking Ultrasound Foundation Models for Fetal Plane Classification

Structure-Augmented Standard Plane Detection with Temporal Aggregation in Blind-Sweep Fetal Ultrasound

Foundation Model-driven Key Anatomy Frame Selection for Blind-sweep Ultrasound Fetal Birth Weight Estimation