Source-linked AI summary
Physics-Guided Synthetic High-Frequency Ultrasound Generation for Skin Layer Segmentation
Junkyung ju, Kyungho Yoon, Minwoo Shin
TL;DR
Dense labels for deeper skin layers are scarce in real HFUS datasets, limiting the scope of supervised skin-layer analysis. This paper generates paired synthetic images, masks, and metadata from multilayer acoustic phantoms using k-Wave simulation, then tests synthetic pretraining with real fine-tuning. The approach achieves performance comparable to real-only training and improves mean Dice/IoU in three of four trainable architectures, though a synthetic-real appearance gap remains.
Problem
Real HFUS datasets commonly annotate epidermis and SLEB, while dense labels for deeper structures are rarely available, and image noise and ambiguous boundaries complicate automated skin analysis.
Source: §1 · p. 2Method
The framework uses multilayer acoustic phantoms and k-Wave simulation to generate paired synthetic HFUS images, dense masks, and metadata for segmentation pretraining.
Source: §1 · p. 2, §3.1 · p. 3Results
Takeaways & Limitations
The generated images encode transferable anatomical and textural cues for real-domain skin-layer segmentation.
Source: §6 · p. 9The simulator does not fully match scanner-acquired HFUS statistics, and real-domain evaluation is limited to epidermis and SLEB because deeper real labels are unavailable.
Source: §6 · p. 9
Abstract
from arXiv · showhide
High-frequency ultrasound (HFUS) enables noninvasive visualization of superficial skin structures, but automated skin-layer analysis is limited by the scarcity of densely annotated data. Existing real HFUS datasets commonly provide annotations for superficial targets such as the epidermis and subepidermal low-echogenic band (SLEB), while dense labels for deeper structures such as dermis, subcutaneous tissue, fascia, and muscle are rarely available. We propose a physics-guided synthetic HFUS generation framework for skin layer segmentation. The framework constructs multilayer acoustic skin phantoms, assigns layer dependent acoustic properties, and uses k-Wave simulation to generate paired synthetic HFUS images, dense layer masks, and simulation metadata. To evaluate whether the generated data provide transferable supervision, we use it for downstream segmentation pretraining and fine-tune the models on real Mendeley HFUS data. Synthetic pretraining followed by real fine-tuning achieved real-domain performance comparable to real-only training and improved mean Dice/IoU in three of four evaluated trainable architectures. These results suggest that physics-guided synthetic HFUS images contain transferable anatomical and textural cues for real-domain skin layer segmentation, although further reduction of the synthetic-real appearance gap is needed to enable greater gains. The code and data are available at: https://github.com/Finn-02/synthetic-hfus-skin-layer-segmentation.
1 Introduction
Dense real HFUS labels are scarce beyond superficial targets, while image noise and ambiguous boundaries complicate automated skin analysis. The paper proposes physics-guided synthesis to provide densely labeled multilayer images and test their transfer to real data.
1 Introduction
- Real HFUS images contain speckle noise, weak layer contrast, attenuation, probe-dependent appearance variation, and ambiguous tissue boundaries.Source: §1 · p. 2
- Public HFUS resources mainly annotate epidermis and SLEB, while dense labels for dermis, subcutaneous tissue, fascia, and muscle are rarely available.Source: §1 · p. 2
- Skin-specific physics-guided generation of densely labeled multilayer HFUS remains underexplored, motivating synthesis that preserves layer order while varying thickness, interfaces, and texture.Prior synthetic ultrasound work supports downstream learning, but the paper distinguishes acoustic layer-map-driven generation from image-level style transfer.Source: §1 · p. 2
- The framework combines multilayer acoustic phantoms and k-Wave simulation to generate dense eight-class labels and tests their transfer through real-domain epidermis/SLEB segmentation.Synthetic pretraining followed by real fine-tuning achieved performance comparable to real-only training and improved mean Dice/IoU in three of four trainable architectures.Source: §1 · p. 2
2 Related Work
Prior HFUS skin studies have focused mainly on real-image analysis, while synthetic ultrasound work and acoustic simulators offer complementary routes to learning and image generation. Physics-guided generation of densely labeled multilayer HFUS skin images remains underexplored.
2 Related Work
- HFUS has been used to assess epidermal thickness, dermal echogenicity, SLEB-related changes, inflammatory dermatoses, skin aging, and aesthetic treatment effects.Source: §2 · p. 2
- HFUS skin studies address classification, quality assessment, and segmentation, but primarily use real images rather than physics-guided generation of dense multilayer datasets.Automated targets include epidermis/SLEB and, in recent work, entry echo, SLEB, and dermis.Source: §2 · p. 3
- Diffusion and GAN ultrasound synthesis can support downstream learning, whereas Field II and k-Wave provide physics-based simulation routes without the same explicit multilayer skin-label focus.The reviewed literature leaves physics-guided generation of densely labeled multilayer HFUS skin images for segmentation pretraining underexplored.Source: §2 · p. 3
3 Method
The method maps multilayer acoustic skin phantoms through k-Wave simulation to paired synthetic images and dense masks, with controlled anatomy and appearance variation. It exports data in a format and task mapping suited to real-domain epidermis/SLEB evaluation.
3.1 Overview of Physics-Guided Synthetic HFUS Generation
- Known anatomical layer maps and layer-dependent acoustic properties are converted by k-Wave into synthetic HFUS images, dense masks, and traceable simulation metadata.The pipeline stores acoustic maps, class maps, visual masks, overlays, and generation metadata for training across the full label space or a task-specific mapping.Source: §3.1 · p. 3, Figure 1 · p. 4
3.2 Multilayer Acoustic Skin Phantom
- The phantom defines eight classes and preserves separate epidermis and SLEB labels, a gradual dermis-to-subcutis transition, thin fascia, and deeper anisotropic muscle texture.SLEB is modeled as a low-scatter inserted band; the figure shows density and sound-speed maps alongside simulated images and the dense layer mask.Source: §3.2 · pp. 3–4, Figure 1 · p. 4
3.3 Structural and Appearance Variation
- Three independent controls vary layer thickness, boundary shape, and tissue heterogeneity, while label-preserving augmentation changes exported image-mask pairs after simulation.This separates anatomy-level variation before image formation from appearance augmentation afterward.Source: §3.3 · p. 5
3.4 k-Wave Simulation, Export, and Task Mapping
- The generator exports 20 MHz simulations at 2067×1555 pixels and maps its eight-class masks to other, epidermis, and SLEB for real-domain comparison.Green-channel rendering is used as input to match the real HFUS display format; the dataset contains 200 original synthetic samples.Source: §3.4 · p. 5
4 Experiments
The experiments generate varied synthetic samples with exact masks, then assess synthetic pretraining followed by real fine-tuning on Mendeley HFUS data. Real-domain evaluation is restricted to epidermis and SLEB because deeper real labels are unavailable.
4.1 Experiment 1: Generating the Synthetic Dataset
- Synthetic samples vary layer thickness, boundary morphology, and tissue heterogeneity, with label-preserving appearance augmentation and dense masks generated from known class maps.Each synthetic image is paired with pixel-level supervision and metadata without manual annotation.Source: §4.1 · pp. 5–6
4.2 Experiment 2: Downstream Segmentation Pretraining
- The transfer experiment compares real-only training against synthetic pretraining followed by real fine-tuning across four trainable architectures, evaluating epidermis and SLEB on Mendeley HFUS.Metrics include class-wise Dice and IoU plus mean Dice and mean IoU; real-domain multilayer accuracy cannot be measured because deeper labels are unavailable.Source: §4.2 · p. 6
5 Results
The synthetic set varies layer structure and image appearance while retaining paired supervision. On Mendeley epidermis/SLEB segmentation, synthetic pretraining followed by real fine-tuning remained competitive with real-only training.
5.1 Experiment 1: Synthetic Dataset Generation
- Synthetic examples vary dermis thickness, boundary irregularity, tissue heterogeneity, and rotation while retaining exact pixel-level correspondence with the mask.Structural variation comes from thickness and boundaries; heterogeneity and rotation provide appearance variation for pretraining.Source: §5.1 · pp. 6–7, Figure 2 · p. 7
5.2 Experiment 2: Downstream Transferability Evaluation
- 83.40% mean Dice and 71.79% mean IoU were Fresh SegUNet’s best values after synthetic pretraining and real fine-tuning; this approach improved mean Dice/IoU in three of four architectures.Across trainable architectures, performance was comparable to real-only training.Source: §5.2 · p. 7
6 Discussion
The framework uses explicit acoustic layer maps to pair synthetic images with dense masks and metadata. Results suggest transferable cues after real fine-tuning, while scanner-appearance differences constrain gains and real evaluation covers only epidermis and SLEB.
6 Discussion
- Explicit acoustic layer maps generate paired synthetic images, dense masks, and metadata, distinguishing the framework from purely data-driven synthesis.The paired outputs connect known layer structure to pixel-level supervision.Source: §6 · p. 7
- After real fine-tuning, synthetic pretraining achieved performance comparable to real-only training and improved mean Dice/IoU in three of four trainable architectures.The authors interpret this as evidence that synthetic images contain layer-structured cues reusable by real-domain models.Source: §6 · p. 9
- The simulator captures transferable geometry and coarse tissue appearance but does not fully match scanner HFUS statistics; more samples may preserve appearance bias, and each physics-based sample is costly.Differences in speckle, attenuation, boundary sharpness, and reflector artifacts may explain architecture-dependent gains.Source: §6 · p. 9
- Although the simulator labels deeper tissues, real-domain evaluation is limited to epidermis and SLEB; reducing the appearance gap is important for extending transfer toward multilayer segmentation.The synthetic labels include dermis, subcutaneous tissue, fascia, and muscle.Source: §6 · p. 9
7 Conclusion
The framework produced structurally diverse synthetic HFUS samples and achieved real-domain segmentation performance comparable to real-only training, supporting its use as an initial validation of transferable synthetic supervision.
7 Conclusion
- Real-domain performance was comparable to real-only training, with improved mean Dice/IoU in three of four trainable architectures.Source: §7 · p. 9
- The generated synthetic HFUS images provided transferable cues for real-domain skin-layer segmentation.Source: §7 · p. 9
- The framework provides a basis for studying multilayer HFUS segmentation under controlled synthetic variation, with real-domain evaluation limited to the clinically annotated epidermis/SLEB task.The authors describe this as an initial validation, not a final solution to real-domain segmentation.Source: §7 · p. 9