PSFH Segmentation Benchmarks Overview

• PSFH segmentation benchmarks are rigorously validated datasets and evaluation protocols for delineating Pubic Symphysis and Fetal Head in challenging ultrasound images.
• They incorporate comprehensive manual annotations, standard splits, and advanced metrics like Dice Similarity Coefficient, HD95, and ASD to ensure a fair cross-method comparison.
• Leading methods leverage hybrid CNN–ViT architectures and innovative techniques such as ellipse constraints and boundary-focused augmentation to improve segmentation under difficult conditions.

PSFH segmentation benchmarks encompass a set of rigorously validated datasets and evaluation protocols dedicated to segmenting anatomically or structurally distinct regions—principally, the Pubic Symphysis (PS) and Fetal Head (FH)—in intrapartum ultrasound or analogous imaging modalities. The benchmarks serve as the de facto standard for quantitative comparison of segmentation algorithms in labor monitoring, semi-supervised medical image analysis, and, in a distinct context, few-shot and point cloud segmentation. The acronym "PSFH" is primarily associated with pixel-wise anatomical segmentation in fetal–maternal ultrasound studies; contextually, it also appears in façade or housing point-cloud corpus annotations. This article surveys the structure and evolution of PSFH segmentation benchmarks, spanning dataset construction, evaluation metrics, methodological trends, and open challenges.

1. Datasets and Task Formulation

The canonical PSFH benchmarks originate from the MICCAI 2023 and 2024 Grand Challenges on Pubic Symphysis and Fetal Head Segmentation (PSFH/PSFHS), providing the most statistically robust and publicly accessible corpus for semi-supervised and supervised evaluation (Bai et al., 2024). The core dataset comprises 5,101 2D B-mode intrapartum ultrasound frames, with comprehensive manual pixel-wise annotations for both the PS and FH regions, collected across multiple institutions (Nanfang Hospital, Zhujiang Hospital, Jinan University) and ultrasound platforms (ObEye, Esaote MyLab). For generalization tests, such as the PSFH MICCAI 2024 benchmark, 300 further images serve as out-of-distribution evaluation. The segmentation target is the delineation of PS and FH masks in images characterized by low contrast, high speckle noise, and boundary ambiguity.

Key dataset characteristics include:

• Standard split: 70% training (≈3,570–4,000 images), 10% validation, 20% testing (≈1,020 images for MICCAI 2023; 700 in-stage 2 testing (Bai et al., 2024)).
• Class distribution: PS is substantially smaller and lower contrast than FH, contributing to class imbalance and boundary detection challenges.
• Annotation protocol: Multi-rater annotation followed by senior adjudication ensures high intra- and inter-observer reliability (PSFH Dice scores 85.2%–90.0%).

Additional PSFH benchmarks, in the context of few-shot semantic segmentation (e.g., PASCAL-5ᶦ, COCO-20ᶦ, FSS-1000 (Catalano et al., 2023, Zhang et al., 2021)) and point cloud façade segmentation (e.g., TUM-FAÇADE (Wysocki et al., 2023)), use the term for structurally analogous multi-class segmentation, though not in obstetric ultrasound.

2. Evaluation Metrics and Protocols

PSFH benchmarks consistently employ rigorous quantitative metrics to facilitate fair cross-methodology assessment:

• Dice Similarity Coefficient (DSC): Measures geometric overlap of predicted versus reference binary masks for each anatomical structure, defined as

$$\mathrm{DSC}(A, B) = \frac{2\,|A \cap B|}{|A| + |B|}$$

• 95% Hausdorff Distance (HD$_{95}$): Captures the 95th percentile boundary discrepancy, mitigating sensitivity to outliers:

$$\mathrm{HD}_{95}(A, B) = \max\left\{ \mathrm{percentile}_{95}(\{d(a, S_B)\}_{a \in S_A}), \mathrm{percentile}_{95}(\{d(b, S_A)\}_{b \in S_B}) \right\}$$

• Average Surface Distance (ASD): Symmetric mean boundary-to-boundary distance,

$$\mathrm{ASD}(A,B) = \frac{1}{|S_A|+|S_B|}\left( \sum_{a\in S_A} \min_{b\in S_B} d(a,b) + \sum_{b\in S_B} \min_{a\in S_A} d(b,a) \right)$$

Ranking in the Grand Challenges is performed over the aggregate of Dice, Jaccard, HD$_{95}$, and ASD across PS, FH, and combined "PSFH" masks. In point cloud or few-shot settings, mean Intersection-over-Union (mIoU), F$_1$-score, and boundary-specific IoU/accuracy are also standard (Wysocki et al., 2023, Catalano et al., 2023).

3. Benchmark Results and Leaderboard Analysis

Performance on PSFH segmentation remains a litmus test for both clinical and methodological advances in medical image segmentation. On the MICCAI 2023 and 2024 test sets, the top methods achieve:

Method/Team	DSC (PSFH)	ASD (mm)	HD$_{95}$ (mm)	Key Backbone
Aloha (2023)	0.927	3.35	13.22	SAM (ViT-h), LoRA fine-tuning
DSTCS (2023)	0.911	0.336	2.070	Dual Student, U-Net + SAM, EPIS augment.
HDC	0.889	0.536	3.673	Single Teacher, Hierarchical Distillation
ERSR	0.937	0.16	1.40	U-Net, Ellipse-constrained pseudo-label

Values above taken from (Bai et al., 2024, Luo et al., 27 Jan 2026, Le et al., 14 Apr 2025, Zhou et al., 27 Aug 2025)

Transformer-based foundation models (notably SAM with LoRA fine-tuning) and dual student–teacher architectures integrating CNN and ViT backbones decisively outperform classic CNN or hybrid U-Net designs (statistically significant, $p<0.05$) (Bai et al., 2024). Models leveraging global anatomical priors (ellipse constraints, symmetry regularization) or advanced augmentation concentrate gains on challenging contours and class-imbalanced regions, notably PS.

Qualitative failure modes for less performant models include PS omission, FH boundary leakage, and over-segmentation in shadow-dominated regions (Bai et al., 2024). Top-ranked ensembles combine architecture variety (CNN, ViT), patch-wise augmentation, and optimal loss mixes (Dice, cross-entropy, focal/Hausdorff).

4. Methodological Advances in PSFH Segmentation

Key methodological themes distinguishing PSFH benchmarks include:

• Dual/multi-student and teacher architectures: DSTCS employs cooperative learning between a CNN (U-Net) and an adapted SAM ViT branch, with an EMA teacher for cross-regularization. This synergistically captures both local texture and global spatial priors, suppressing pseudo-label noise and constraining ambiguity at class boundaries (Luo et al., 27 Jan 2026).
• Advanced supervision and consistency objectives: HDC introduces hierarchical distillation: Correlation Guidance Loss tightly aligns student–teacher feature maps, and Mutual Information Loss regulates student stability under stochastic perturbations (Le et al., 14 Apr 2025). ERSR imposes ellipse-constrained pseudo-label refinement and symmetry-based consistency regularization, encoding robust anatomical priors (Zhou et al., 27 Aug 2025).
• Boundary-focused augmentation and losses: Techniques such as Edge-Patch In-Situ Superposition (EPIS) inject boundary-centric patches, and Neighborhood Weighted Dice Loss upweights optimization on pixels with high neighbor disagreement, further sharpening delineations in ambiguous regions (Luo et al., 27 Jan 2026).
• Robust generalization: Across MICCAI 2024 and out-of-distribution test sets, top methods sustain high DSC and low ASD, exhibiting minimal missed cases or outliers—attributed to the integration of large pre-trained encoders, strong augmentation, and distribution-aware supervision (Luo et al., 27 Jan 2026).

5. Extensions to Few-Shot, Point Cloud, and Text Segmentation

The term "PSFH segmentation benchmark" also arises in structurally cognate but domain-specific contexts:

• Few-shot semantic segmentation: Benchmarks such as PASCAL-5ᶦ and COCO-20ᶦ define episodic protocols (1-shot, 5-shot; novel class generalization) with mIoU and FB-IoU as reporting metrics. Notable approaches include meta-prototype hierarchies and prior-enhanced foreground extraction, leading to measurable improvements in IoU, especially in boundary regions (Zhang et al., 2021, Catalano et al., 2023).
• Point cloud façade segmentation: The TUM-FAÇADE benchmark defines a standardized workflow for enriching urban MLS point clouds with a 17-class hierarchical annotation schema and provides strict mIoU, per-class F$_1$, and OA metrics. This procedural protocol enables scalable, reproducible ground-truthing of façade elements in city-scale datasets (Wysocki et al., 2023).
• Text and speech segmentation: In a distinct setting, the PSFH (Paragraph Segmentation of Formal and Heterogeneous speech) benchmarks, TEDPara and YTSEGPara, codify paragraph boundary detection in transcripts using sentence-aligned constrained decoding, with F$_1$, P$_k$, and WindowDiff as core metrics. These benchmarks structure unformatted transcripts for readability and downstream NLP tasks (Retkowski et al., 30 Dec 2025).

6. Best Practices, Limitations, and Recommendations

Benchmarking in PSFH segmentation has fostered several best practices and elucidated enduring challenges:

• Best practices: Employ transformer-based or hybrid CNN–ViT architectures; leverage strong augmentation pipelines; ensemble models for outlier suppression; calibrate composite loss functions incorporating geometric and pixel-wise terms; validate cross-institutional and technical diversity to ensure generalizability (Bai et al., 2024, Luo et al., 27 Jan 2026).
• Dataset design: Ensure diversity in device, operator, patient population, and protocol; document annotation strategy and inter-observer agreement; provide publicly accessible training and benchmarking toolkits (Bai et al., 2024).
• Open challenges: Persistent sources of error include PS class imbalance, device/operator-specific artifacts, and boundary uncertainty under severe acoustic shadowing or deformations. Over-regularization by strong anatomical priors (e.g., ellipses) may impair generality in non-elliptical structures (Zhou et al., 27 Aug 2025).
• Recommendations: Advance towards video-based, spatio-temporal segmentation; integrate explicit anatomical or shape priors; pursue lightweight distillation of foundation models for deployment; expand dataset diversity (geography, devices, annotation granularity) and standardize cross-domain evaluation pipelines (Bai et al., 2024, Luo et al., 27 Jan 2026).

7. Public Resources and Community Impact

PSFH segmentation benchmarks have established a reproducible, openly accessible basis for algorithm development, comparison, and future research. The MICCAI 2023/2024 PSFH datasets (training, test, challenge platform), code for top entries, and evaluation toolkits are publicly available (Bai et al., 2024), catalyzing broader progress in obstetric ultrasound analysis and beyond. Leaderboards from these challenges directly influence methodological trends: all top-performing submissions are hybrid or transformer-based, validating the shift to large-scale, pre-trained, and anatomically informed architectures. These benchmarks underpin further advances in semi-supervised learning, anatomical prior integration, and domain-adaptive segmentation for clinical translation.