Source-linked AI summary
Anatomy-Guided Foundation Model Adaptation with Within-Case Prototype Supervision for Standard Plane Detection in Fetal Ultrasound Blind Sweeps
Yuzhe Zhao
TL;DR
AnatoProto addresses standard-plane detection in fetal ultrasound blind sweeps, where positive frames are rare, short, and poorly handled by existing foundation models. It adapts frozen BiomedCLIP with anatomy and case-level structure, achieving F1 = 67.72 and revealing that accurate anatomy-guided prototypes make prototype supervision effective.
Problem
Fetal abdominal circumference standard-plane detection in low-cost blind sweeps requires locating rare, short positive segments from non-expert-acquired videos.
Method
AnatoProto adapts a frozen BiomedCLIP encoder using anatomy-weighted pooling, within-case prototype supervision, cascade refinement, and a hybrid stability–boundary head.
Results
F1 = 67.72, outperforming FetalCLIP + PRS by +13.20 F1 and lifting performance from 58.37 to 67.72 over the baseline.
Takeaways & Limitations
The anatomy prior makes within-case positive prototypes anatomically accurate, turning prototype loss from a liability into a lever and producing a non-additive synergy.
Takeaways & Limitations
Performance remains constrained by very short ground-truth planes and by the coarse, task-specific, and potentially unreliable nnU-Net anatomy prior.
Abstract
from arXiv · showhide
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.
1 Introduction
AC standard-plane detection in low-cost fetal blind sweeps is dominated by rare, contiguous positives, while frame-level foundation models and generic temporal detectors underperform. AnatoProto addresses this with anatomy-guided pooling, within-case prototypes, cascade refinement, and hybrid temporal prediction, reaching F1 = 67.72.
- Problem: Fewer than 3% of frames are positive and positives form short contiguous segments, requiring sweep-level rather than frame-level reasoning.The task labels each frame in an 840-frame blind sweep as optimal, suboptimal, or background relative to the AC standard plane.
- Motivation: Single-frame foundation models and off-the-shelf temporal action detectors saturate below approximately F1 = 55 and F1 ≈52, respectively.The reported comparison attributes the gap to missing temporal and structural task-specific reasoning rather than model capacity.
- Approach: AnatoProto adapts a frozen BiomedCLIP encoder with anatomy-weighted spatial pooling, a within-case prototype loss, cascade refinement, and hybrid stability–boundary prediction.The anatomy prior reweights BiomedCLIP patch tokens, while the other components exploit sweep-level structure and boundary behavior.
- Approach: The anatomy-weighted pool uses nnU-Net abdominal-region probabilities to steer frozen BiomedCLIP features toward anatomically meaningful regions.In isolation, the pool raises F1 by +4.21 over the baseline.
- Approach: The prototype loss reduces recall by 12 points alone but increases recall by +6.5 points when combined with anatomy-weighted pooling.The reported sign flip is linked to whether the within-case positive prototype is accurate.
- Results: F1 = 67.72, improving by +13.20 over FetalCLIP + PRS at 54.52 and by +30.33 over a BiomedCLIP linear probe at 37.39.Evaluation uses a held-out case-level ACOUSLIC-AI split, with paired-bootstrap significance testing reported for performance gaps.
2 Related Work
Prior work spans curated-clip frame classifiers, ultrasound and medical foundation models, temporal action detectors, and mask-guided representation learning. These approaches expose the difficulty of blind sweeps, where rare short segments and missing anatomical or case-level priors limit performance.
- Standard-plane detection: Curated-clip frame classifiers and the ACOUSLIC-AI segmentation baseline do not translate well to blind sweeps with rare, background-dominated positives.The ACOUSLIC-AI baseline plateaus at F1 = 13.78, with Precision = 10.41 and Recall = 20.39.
- Foundation models: Ultrasound and medical foundation models show domain, input, or imbalance limitations, while frozen FetalCLIP with PRS reaches F1 = 54.52 on isolated frames.The cited limitations include breast-versus-fetal domain mismatch, reliance on unavailable masks, grayscale ultrasound shift, and collapse under a 2.6% positive rate.
- Temporal action detection: ActionFormer and TriDet with PRS saturate at F1 ≈52 because target segments are short, highly imbalanced, and visually near-duplicate.These generic detectors lack mechanisms for anatomical and within-case structural priors.
- Mask-guided learning: Mask-guided representation learning motivates steering pretrained visual encoders with external spatial priors and coupling segmentation with downstream anatomical tasks.AnatoProto adopts this general direction through anatomy-weighted pooling.
3 Method
AnatoProto adapts frozen BiomedCLIP features for fetal blind-sweep standard-plane detection by combining anatomy-aware pooling, within-case supervision, temporal sequence modelling, and coarse-to-fine inference. The method converts noisy frame predictions into structurally constrained segment decisions while addressing severe class imbalance and short contiguous positives.
- Problem setting: Positive frames form fewer than 3% of each sweep and occur in short contiguous segments, requiring sweep-level rather than frame-level reasoning.The method merges optimal and suboptimal AC standard-plane frames into one positive class.
- Architecture: AnatoProto uses frozen BiomedCLIP patch tokens, an nnU-Net abdominal probability map, and a 4-layer BiTT to produce frame-level sequence predictions.The BiTT jointly supports frame, stability, and boundary heads under a hybrid training objective.
- Anatomy-weighted spatial pooling: Anatomy-weighted pooling reweights semantic patch tokens toward the fetal abdomen without adding learnable parameters.The coarse 7 × 7 nnU-Net map acts as a spatial prior, while BiomedCLIP supplies the semantic patch descriptions.
- Within-case prototype loss: The within-case prototype loss averages positive-frame embeddings within each sweep and pulls positives toward that prototype while pushing negatives away.The loss exploits physical similarity among adjacent positive frames, which frame-level BCE does not encode.
- Coarse-to-fine inference: The cascade lifts predictions from frames to segments using PRS, then rejects implausible segments with a logistic regression over stability, boundary, length, and position features.PRS merges gaps of at most 2 frames and removes segments shorter than 5 frames; the rejecter uses eight scalar features.
4 Experiments
On a held-out 45-case ACOUSLIC-AI test split, AnatoProto substantially outperforms foundation-model and temporal-action baselines. Ablations and qualitative analyses attribute gains to anatomy-guided prototypes, sequence-aware refinement, and complementary stability and boundary modeling.
- Comparison with State of the Art: F1 = 67.72, exceeding FetalCLIP + linear probe + PRS by +13.20 F1 and TriDet + PRS by +15.76 F1.The D-3 versus FetalCLIP+PRS gap is statistically significant under paired bootstrap, while its gap versus the uniform-pool baseline is not.
- Comparison with State of the Art: AnatoProto reaches Precision 63.58 / Recall 72.44, improving the trade-off between FetalCLIP’s high Recall and temporal detectors’ lower Recall.FetalCLIP + PRS attains Precision 39.08 / Recall 90.16, while ActionFormer and TriDet show comparatively higher Precision but lower Recall.
- Ablation: Anatomy Pool × Prototype Loss: +6.5 Recall points arise when prototype loss is added to anatomy-weighted pooling, whereas the same loss reduces Recall by 12.0 points with uniform pooling.Both marginal contrasts are statistically significant under 2,000-run paired bootstrap; the corresponding F1 synergy is not significant.
- Ablation: Anatomy Pool × Prototype Loss: Anatomy-weighted pooling anchors prototypes on fetal anatomy, while uniform pooling produces contaminated prototypes that can become confidently wrong attractors.Before training, anatomy weighting raises margin_sat+ from 0.537 to 0.645 and between-class prototype distance from 0.257 to 0.308.
5 Discussion and Conclusion
AnatoProto combines anatomy and case-level structure to improve standard-plane detection in fetal ultrasound blind sweeps, while exposing limitations around short segments, generalisation conditions, and spatial-prior reliability.
- F1 rises from 58.37 to 67.72, outperforming the strongest external baseline by +13.20 F1 at a matched precision/recall trade-off.
- The anatomy prior and prototype loss interact synergistically: accurate abdominal prototypes convert prototype supervision from a liability into a recall-improving mechanism.
- Failure-mode analysis: AnatoProto still loses recall when true planes span only 2–3 frames, because the PRS threshold and 5-frame minimum-length filter discard short segments.
- Generalisation: The method’s transferable ingredients require a trainable anatomical mask and within-recording repetition of positive frames as short contiguous plateaus.
- Generalisation: When these conditions weaken, the anatomy-weighted pool may lose most of its advantage and the prototype loss may require re-thinking.
- Clinical implications: At F1 = 67.72, AnatoProto achieves Precision 63.58 and Recall 72.44, while reducing reviewer frame-scan burden by roughly 4.9× versus the official baseline in triage.
- Limitations and future work: The anatomy prior is coarse and dependent on nnU-Net reliability, while the F1-level synergy is not statistically significant on only 45 held-out cases.