Skin, as the body’s largest organ, serves to regulate body fluid and temperature and forms a protective barrier shielding the organism against pathogens and injuries from the environment (Proksch et al., 2008). Skin disease is one of the most common human illnesses, affecting nearly 900 million people, more than one-third of the global population (Hay et al., 2014). Among these, acne vulgaris is a prevalent chronic skin inflammatory disease affecting up to 85% of late adolescents (Lynn et al., 2016), resulting in various consequences, including scarring, dyspigmentation, and psychological impacts (Ogé et al., 2019). However, there is currently no universally accepted assessment system for acne vulgaris.

The biomechanical properties of skin are primarily determined by its structural components (Joodaki and Panzer, 2018). Elastography is the functional modality to provide information on the biomechanical properties of tissues. Among different elastography modalities, optical coherence elastography (OCE), derived from optical coherence tomography (OCT), has an ultra-fast sampling rate, micrometer imaging resolutions and millimeter depth penetration (∼one to two mm) (Larin and Sampson, 2017). A notable branch of OCE technology is wave-based OCE, an in situ non-destructive approach that quantitatively estimates biomechanical properties in soft tissues using elastic waves (Liang and Boppart, 2009). Biomechanical properties, especially elasticity (Everett and Sommers, 2013), have been proven to be a potential biomarker for characterizing skin aging (Couturaud et al., 1995), understanding physiology, pathological cases, and monitoring treatment (Balbir-Gurman et al., 2002; Neto et al., 2013; Killaars et al., 2015). In OCE, wave propagation in tissue occurs when an elastic wave is generated by excitation and then transmits through other regions of the tissue. The velocity of the wave is intrinsically related to the biomechanical properties of the tissues (Kirby et al., 2017). OCE’s millimeter penetration depth confines motion measurements to regions near tissue boundaries, where surface acoustic waves (SAWs) are the dominant wave type (Zvietcovich and Larin, 2022). SAW velocities can be estimated by analyzing the phase term of the complex OCT signal. Typically, the phase difference between successive scans is utilized to detect sub-resolution axial differential displacement within a sample (Song et al., 2013), followed by the use of a time-of-flight approach to measure SAW velocities. By selecting an appropriate elasticity model, the biomechanical properties of the tissue can then be determined (Zvietcovich and Larin, 2022). While wave-based OCE has gained increasing interest in recent years, its application to in vivo skin conditions remains in its early stages. Two pre-clinical studies have shown the ability of wave-based OCE to characterize mechanical properties in animal models of systemic sclerosis (Du et al., 2016) and skin burns (Liu et al., 2024). However, only one wave-based OCE system has been translated to a clinical trial in human subjects for the assessment of systemic sclerosis in vivo (Liu et al., 2019). The major challenges limiting the clinical translation of OCE technology are the high level of expertise required and the inability to produce real-time results (Sun et al., 2011). In particular, biomechanical property analysis often demands complex image processing for wave feature extraction and velocity estimation (Song et al., 2015; Kirby et al., 2019), which could extend processing times to potentially several minutes or longer, limiting its use in real-time clinical settings.

Deep learning holds considerable promise for enhancing the efficiency of the processing of wave-based OCE by discerning and analyzing raw data. Currently, deep learning-assisted OCE analysis is still in the early stages. Schlaefer’s group (Neidhardt et al., 2020; Neidhardt et al., 2021; Neidhardt et al., 2023) demonstrated elastic velocity prediction for OCE data by using convolutional neural networks (CNNs) with dense connections. These methods have been proved based on homogeneous tissue-mimicking materials (Neidhardt et al., 2020; Neidhardt et al., 2021) and ex vivo chicken heart (Neidhardt et al., 2023) studies. However, there are inherent differences in structural (Labroo et al., 2021) and physical (Godin and Touitou, 2007) properties between heterogeneous animal and human tissue. Additionally, involuntary movements (Kirkpatrick et al., 2006) and breathing motion artefacts (Fang et al., 2019) frequently occur during in vivo human OCE acquisitions. Consequently, their CNN models might need to adapt the intricate textures of wave patterns from in vivo human data, instead of focusing on velocity prediction, leading to less optimal for in vivo human applications.

In this study, we propose a novel velocity prediction network (VP-Net), that predicts bulk (body) wave velocities in vivo human healthy and abnormal skin sites from raw OCE data. The network architecture incorporates a squeeze-and-excitation (SE) block (Hu et al., 2018) and a separable convolution block, enabling efficient feature reuse and integration without significantly increasing model complexity. Compared to existing CNN models, VP-Net could accurately predict elastic wave velocity from each raw phase slice directly, maintaining the lowest model complexity and inference time. VP-Net demonstrated high accuracy in predicting elastic wave velocities in multiple healthy skin sites and distinguishing age-related velocity changes between 20s and 30s age groups. Closed comedones, a type of acne lesions (Lavers, 2014), were also investigated in this study. VP-Net’s successfully predicted high velocities in comedones, indicating elevated skin elasticity. To the best of our knowledge, this is the first study to quantify the biomechanical properties of facial acne lesions using OCE technology and to develop an elastic wave velocity prediction model in human in vivo using deep learning. VP-Net achieved a processing speed of 1.32 ms per slice, approximately 100 times faster than a conventional velocity estimation method. Therefore, VP-Net offers real-time elastic wave velocity prediction in human skin in vivo, providing potential clinical applications in characterizing skin aging, as well as assessing and managing the treatment of acne vulgaris.

Our study has five main contributions: 1) Our model demonstrated consistent and repeatable velocity predictions on tissue-mimicking phantoms, which are homogenous and have consistent biomechanical properties for each concentration. 2) To the best of our knowledge, this is the first study to deploy a deep learning method to directly predict biomechanical property-related velocities for in vivo human healthy and abnormal datasets, showcasing its potential for skin condition diagnosis. 3) We conducted a comprehensive comparison with various neural networks and an ablation study on VP-Net to validate the efficacy of our proposed model. 4) Compared to existing models, the proposed VP-Net has the fastest inference time and the lowest model complexity while providing accurate SAW velocity predictions, even when applications shifted from tissue-mimicking materials to in vivo human skin. 5) We used gradient-based class activation maps (Grad-CAM) to visualize the model’s process in predicting velocities.

This paper is structured as follows: The Methods section describes the details of our proposed velocity prediction deep learning model and the OCE data processing strategies to generate raw phase slices and ground truth velocities. The Results section presents the performance metrics, ablation study, and visual explanations of our network’s efficacy in predicting velocities. Additionally, the predicted bulk velocities of agar phantoms and healthy skin sites from participants across two age groups, as well as abnormal skin, are shown. Finally, we conclude the paper with a summary of our key contributions, a discussion on the factors affecting model performance, and potential improvements for future research.

Wave-based OCE has been one of the most studied OCE branches, producing a fundamental impact in the quantitative and nondestructive biomechanical characterization of tissues. However, the long processing time limits its real-time and clinical applications (Sun et al., 2011). In this study, we proposed a rapid, high-efficiency, and high-accuracy deep-learning-based velocity prediction network (VP-Net) to predict biomechanical property-related velocity. We comprehensively evaluated the network with homogenous agar-based tissue-mimicking phantoms, in vivo human healthy and abnormal skin. Compared to the conventional OCE velocity estimation method (Zvietcovich and Larin, 2022), VP-Net could directly predict velocity from a single raw OCE slice, which provided end-to-end processing and eliminates the requirement for complex processing. Therefore, the proposed VP-Net has great potential to be translated into clinical practice for characterizing skin aging, as well as assessing and managing the treatment of acne vulgaris.

In the discussion, the results will be analyzed and compared with the findings from other studies. First, we conducted a comprehensive comparison with a series of existing deep-learning models, including VGG16/19, ResNet18/34/50/101, DenseNet121/169, and MobileNetV2. The evaluation results in Table 1 and Table 2 show that the mean MSE and MAE errors were approximately below 0.5 in agar phantoms, with concentrations ranging from 1.5% to 4%, indicating high accuracy in predicting the velocities in these agar phantoms. However, for the agar phantoms with low (1%) and higher concentrations (4.5% and 5%), the mean errors from VP-Net were relatively higher than 0.5. We hypothesize this is due to the unbalanced data distribution in the training datasets, as the velocity distributions for these concentrations had fewer slices (5,127 slices). Regarding the in vivo human skin (Table 3), VP-Net achieved the best performance in the back of hand (MSE: 1.585; MAE: 0.992) and had the lowest MAE of 0.863 in the forearm. In the palm, VP-Net performed similarly to ResNet101 in terms of MSE and MAE. For closed comedones, VP-Net had the second-lowest MSE and MAE. Thus, VP-Net demonstrated high accuracy in predicting biomechanical property-related velocities, indicating its potential for early diagnosis of skin conditions.

An ablation study was conducted to investigate the influence of VP-Net sizes on performance. As shown in Table 4, increasing the size of VP-Net did not improve accuracy for agar phantoms with 1.5%–4.0% concentrations. However, for agar phantoms with 1.5%–2.5% concentrations, decreasing the size of VP-Net improved performance. In the human skin dataset (Table 5), VP-Net-B provided the lowest MAE (0.992) and MSE (1.585) errors in the back of hand, and the lowest MAE (0.997) and second-lowest MSE (2.007) in the forearm. In the palm and closed comedones, reducing the size of VP-Net again provided the best performance in terms of MSE and MAE errors. Compared to VP-Net-S and VP-Net-L, VP-Net-B offered the best trade-off between prediction performance and model complexity.

Additionally, we evaluated the computational demand of VP-Net in both GPU and CPU environments, comparing inference time and model complexity among various methods, as presented in Figure 4. Figure 4A, B illustrate that VP-Net had the lowest inference time in both environments. Specifically, Figure 4A shows that VP-Net accelerated the velocity prediction procedure by a factor of 100 compared to the conventional method. Figure 4B further indicates that VP-Net-S and VP-Net-B had the lowest model complexity and network parameters, respectively.

Grad-CAM (Figure 5) was employed to interpret VP-Net’s velocity prediction processes. When the wave propagation pattern was clear and had single wave mode details (Figure 5 (1B and 2B)), the full wave propagation path (Figure 5 (1C,D and 2 C,D)), was seen in the shallow convolution layers (first to third). In contrast, when artifacts induced by motion, far-end noise, or low intensities at deeper depths were present (Figures 5, 3B, 4B), only the high-quality portions of the wave patterns were emphasized in these layers (Figures 5, 3C, D). This may indicate that the model effectively filtered significant noise from the raw phase slices to extract useful and accurate wave information. For abnormal skin, only the high-velocity wave propagating through the comedo region was displayed as the highest intensity curve in the first convolution layer (Figure 5C). We believe that the comprehensive training dataset, which included high-quality slices at surface depths, low-intensity wave images at deeper depths, motion artifacts, and boundaries between abnormal and healthy regions, enhanced the model’s ability to analyze difficult situations and accurately predict the velocity of abnormal regions (Li et al., 2022).

Biomechanical properties, specifically elasticity (Young’s modulus), can be estimated directly from velocity measurements. The bulk Young’s modulus ( E ) can be calculated from the predicted SAW velocity ( V R ^ ) using the formula E = 3.35 × V R ^ 2 , assuming skin mass density (ρ) of 1.02 g/cm² and a Poisson’s ratio (υ) of 0.5 (Liang and Boppart, 2009). By converting the velocities in Table 6 to elasticity, the bulk Young’s modulus for 1%–2% agar phantoms predicted from raw phase slices were 44 ± 14 kPa, 66 ± 4 kPa, and 149 ± 14 kPa, respectively. These results are in good agreement with the values obtained by Yang et al. (Yang et al., 2022). The Young’s modulus for the 2% agar phantom predicted by our model was 286 ± 12 kPa, comparable to the 254 kPa reported by Brewin et al. (Brewin et al., 2015). In our study, the average predicted Young’s moduli for the three skin sites were 188 ± 58 kPa for the palm, 79 ± 42 kPa for the forearm, and 64 ± 35 kPa for the back of hand, was consistent with values documented in our previous study (Zhou et al., 2020). Also, the values aligned with other reference values. For instance, The bulk Young’s modulus for the palm was reported at 108 ± 48 kPa (Zhang et al., 2011), forearm at 42 ± 32 kPa (Zhang et al., 2011) and 129 ± 88 kPa (Diridollou et al., 2000); Back of hand 11–23 kPa (Wakhlu et al., 2017). Notably, a difference in bulk Young’s modulus between age groups (20s and 30s) was observed: for the palm, 154 ± 37 kPa in the 20s group vs 220 ± 40 kPa in the 30s group; for the forearm, 62 ± 9 kPa in the 20s group vs 95 ± 38 kPa in the 30s group; and for the back of the hand, 65 ± 9 kPa in the 20s group vs 69 ± 7 kPa in the 30s group. Regarding abnormal skin, no previous studies have reported the Young’s modulus of facial skin diseases. For comparison, the mean stiffness of malignant neck tumors was 226.4 kPa as measured by ultrasound elastography (Roldán, 2016), similar to the predicted Young’s modulus of closed comedones at 308 ± 35 kPa. Thus, our proposed VP-Net demonstrated its efficacy by accurately obtaining bulk velocity from a single image with noisy raw phase information.

Neidhardt et al. (2021) reported a densely connected network for predicting concentrations of gelatin phantoms by analyzing shear wave OCE data. They later expanded this approach to aid in force estimation on gelatin phantoms and ex vivo chicken hearts (Neidhardt et al., 2023). Their model could process both 3D (depth × lateral distance × time) and 4D (depth × lateral distance × vertical distance × time) volumes, with each dimension of 32 pixels, and was capable of performing classification in real-time. While their methods had valuable contributions, particularly for real-time and 4D analysis, there may be challenges when applying this approach to in vivo studies and clinical translations. First, the input depth for each volume in their model required 32 pixels, approximately 235 µm. In contrast, our proposed deep learning network could predict velocity from each single slice, with a single depth layer of approximately 4.7 µm. In addition, their low spatial sampling points limited the spatial resolution of the raw volume, resulting in reduced elastography resolution (Kirby et al., 2019). This constrains its applications to address motion artifacts and complicated wave patterns, which frequently occur in vivo OCE acquisitions. Next, its field of view was restricted to 3 mm, which could be insufficient for measuring abnormal skin conditions, typically around 6 mm in diameter (Kasmi and Mokrani, 2016). In our study, the scanning range was up to 11 mm, and we successfully predicted the bulk velocities of closed comedones with diameters greater than 4.2 mm. Thus, VP-Net may offer an advantage in predicting biomechanical property-related velocity from a single image, handling high noise and artifacts, and is particularly suitable for both healthy and abnormal in vivo scans.

While our work represents a significant advancement, further research is needed to refine the deep learning model, particularly its translation to clinical settings. By including a more diverse range of participants, we intend to enhance the robustness of our model, ensuring accurate wave velocity predictions across all biological genders. Additionally, we plan to substantially enlarge our dataset to explore the potential of vision transformers for predicting the biomechanical properties of both healthy and abnormal human skin.

In conclusion, we developed an end-to-end deep learning-based velocity prediction network (VP-Net) for predicting elastic wave velocities associated with biomechanical properties using OCE. VP-Net demonstrated the ability to provide real-time elastic wave velocity predictions without the need for expertise and complex image processing. In vivo applications on both healthy and abnormal human skin, VP-Net accurately differentiated age-related changes in elastic velocities across multiple skin sites and detected high velocities in closed comedones. Therefore, VP-Net holds significant potential for clinical applications in characterizing skin aging, as well as assessing and managing the treatment of acne vulgaris.