VP-net: an end-to-end deep learning network for elastic wave velocity prediction in human skin in vivo using optical coherence elastography

Yilong Zhang, Jinpeng Liao, Zhengshuyi Feng, Wenyue Yang, Alessandro Perelli, Zhiqiong Wang, Chunhui Li, Zhihong HuangFrontiers in Bioengineering and Biotechnology 2024 · DOI

Abstract

This study indicates that VP-Net could rapidly and accurately predict elastic wave velocities related to biomechanical properties of in vivo healthy and abnormal skin, offering potential clinical applications in characterizing skin aging, as well as assessing and managing the treatment of acne vulgaris.

Introduction

The biomechanical properties of skin, such as elasticity and stiffness, are critical indicators of health, aging, and disease. Wave-based Optical Coherence Elastography (OCE) is a non-invasive technique that measures these properties by tracking the speed of induced elastic waves—specifically Surface Acoustic Waves (SAWs)—through tissue. Generally, faster wave propagation indicates stiffer tissue.

However, the current "gold standard" for estimating this velocity, the Time-of-Flight (ToF) method, faces significant clinical hurdles. It requires substantial expertise, complex noise filtering, and lengthy processing times that can stretch into minutes. Furthermore, in vivo measurements are frequently corrupted by involuntary patient movement and breathing artifacts. VP-net was developed to transform OCE from a post-processing-heavy research tool into a rapid, automated clinical diagnostic aid.

Methods

The researchers proposed VP-net, an end-to-end convolutional neural network (CNN) that treats velocity estimation as a linear regression problem. Unlike traditional methods that require manual feature extraction and multi-stage filtering, the network takes raw 2D phase slices (representing spatial-temporal data) as input and directly outputs a single bulk velocity value.

Schematic of the deep learning-based OCE velocity prediction pipeline.

Architectural Innovations

VP-net incorporates several key components to balance speed and accuracy:

  • CBR Blocks: Standard Convolution-Batch Normalization-ReLU layers for initial feature extraction.
  • Separable Convolution Blocks: These reduce model complexity and computational demand by splitting spatial and channel-wise convolutions, enabling high inference speeds.
  • Squeeze-and-Excitation (SE) Blocks: These blocks "recalibrate" feature maps, allowing the network to focus on relevant wave propagation patterns while suppressing background noise and motion artifacts.
Architecture of the Squeeze-and-Excitation (SE) Block used in VP-net.

Datasets

The study utilized a diverse range of data to ensure generalizability:

  • Tissue-Mimicking Phantoms: 16,424 raw phase slices from agar-based phantoms (1% to 5% concentration) provided a ground-truth baseline.
  • Healthy Human Skin: 28,270 slices from 16 participants across three sites (palm, forearm, and back of the hand), including two age groups (20s and 30s) to test sensitivity to aging.
  • Abnormal Human Skin: 580 slices targeting facial closed comedones (acne lesions) to evaluate clinical diagnostic potential.

Results

VP-net demonstrated significant improvements over traditional methods and established deep learning architectures:

  • Speed: The model achieved an inference time of 1.32 ms per slice, representing a 100x speedup over conventional ToF methods.
  • Accuracy and Consistency: VP-net showed lower Mean Squared Error (MSE) and Mean Absolute Error (MAE) than ResNet and DenseNet. Predicted velocities were highly consistent with ToF ground truth, accurately capturing the high stiffness of the palm (6.7–8.3 m/s) compared to the softer forearm (4.3–5.4 m/s).
  • Biomechanical Sensitivity: The model successfully detected that SAW velocity increases with age and is significantly higher in acne lesions (~9 m/s) compared to healthy skin (~4-8 m/s), confirming its utility as a biomarker tool.

Discussion and Conclusion

VP-net offers a robust solution for real-time biomechanical assessment of skin. By mapping raw phase data directly to velocity, it eliminates the need for manual region-of-interest selection and complex pre-processing. While the current results are promising, future work will focus on expanding the dataset to include more diverse skin types and exploring Vision Transformers (ViTs) to further improve performance. The ultimate goal is the integration of VP-net into handheld OCE probes for real-time clinical use in dermatology to monitor conditions like acne or skin aging.

Summary updated

Abstract

This study indicates that VP-Net could rapidly and accurately predict elastic wave velocities related to biomechanical properties of <i>in vivo</i> healthy and abnormal skin, offering potential clinical applications in characterizing skin aging, as well as assessing and managing the treatment of acne vulgaris.
BibTeX
@inproceedings{zhang2024vp,
  author = {Yilong Zhang and Jinpeng Liao and Zhengshuyi Feng and Wenyue Yang and Alessandro Perelli and Zhiqiong Wang and Chunhui Li and Zhihong Huang},
  title = {{VP-net: an end-to-end deep learning network for elastic wave velocity prediction in human skin in vivo using optical coherence elastography}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2024},
  doi = {10.3389/fbioe.2024.1465823},
}
Figure