Non-invasive estimation of skin thickness from hyperspectral imaging and validation using echography
Introduction
The study of human skin and its related properties has now been an active area of research for many decades. There is a plurality of work that ranges from biophysical modeling of skin, to non-invasive diagnostics of skin related conditions [9], [10]. Applications of skin studies span from cancer research, pharmaceuticals, drug delivery, and cosmetics [1]. In addition, there is a significant research interest in designing non-invasive methods for rapidly and accurately estimating the components and properties of human skin. Various research groups have suggested that longitudinal studies of skin components are vital to not only characterizing etiology, and understanding pathophysiology, but also detecting the onset of cancerous tumors [7].
In this paper we discuss a generalized method that is capable of estimating the underlying parameters of human skin. One of these parameters, the skin thickness, is very important in understanding a number of pathologies. Unfortunately, the current method for estimating thickness requires one to excise a sample of skin through a biopsy and then study it under a microscope [8]. This method is not ideal when it needs to be repeated for multiple sites, and possibly multiple sessions. Non-invasively estimating skin thickness is very challenging because skin varies a great deal from one anatomical location to another. It also varies based on ethnicity, environmental factors, and age, amongst others. While many researchers have proposed methods for estimating skin thickness, they seldom have access to any gold standard. In such cases the best that can be done is to assess the physiological plausibility of estimated parameters. Such a benchmarking methodology cannot truly assess the accuracy of the estimation.
Our own prior work includes [16], [17], [18], [19], where we used support vector machines coupled with a Kubelka–Munk (K-M) based biophysical models of human skin. This method was used to estimate melanosomes, collagen, blood volume (each as percentage by volume), and oxygen saturation (as a percentage). It, however, assumed that the thickness of skin was constant, and was hence not estimated. Our biophysical model was based on work by Nunez [12] who developed a model which related the components of skin to a multi-band reflectance spectra. Prior work in this area also includes that of Yudovsky and Pilon [20] who developed a two-layer model and an inverse method based on radiative transfer equations to estimate melanin, epidermal thickness, and oxygen saturation. We differ from that work in that we go beyond measuring epidermal thickness and estimate the total thickness as well as all other skin parameters. In addition, Pilon et al. compared their skin thickness estimation to tabulated values from literature. This is appropriate to determine if the estimates are physiologically plausible, but is insufficient to judge the accuracy of the method because the true estimates for the patients are not available. In that regard, in this study we have access to physician-made gold standard, and hence can directly compute the accuracy of our method. Finally, Claridge and Preece [6], [13] developed a method based on optimizing a set of filters that map digital images and histological parameters. That work estimated the dermal thickness, however it measured the error between the RGB and the optimal filters (not raw parameters). For additional reviews of past work in this area, the reader is referred to our prior study in [16], and to other thorough reviews of K–M models by Baranoski et al. [2], and a review of optical properties of skin by Cheong et al. [5].
The novel features of the study presented herein includes
- (1)
We extend our previous biophysical model to estimate the thickness for 5 strata and 4 dermises of human skin.
- (2)
We acquire Ultrasound (US) sequences corresponding to each hyperspectral signature, from which we derive physician-generated gold standard for the skin thicknesses.
- (3)
We perform a sensitivity analysis with regard to the number of bands used.
Section snippets
Biophysical model of human skin
In this study we extend our previous biophysical model [18], [19] and the one developed by Nunez [12]. This model has 10 layers: layers 1–5 represent the strata of human skin; these layers are: stratum corneum, stratum lucidum, stratum granulosum, stratum spinosum, and stratum basale. Layers 6–9 are the dermises; these layers are: papillary dermis, upper blood net dermis, reticular dermis, and deep blood net dermis. Layer 10 is the subcutaneous tissue layer. As described before [18], [19] it is
Results and discussions
The methods were tested on a separate testing dataset. This dataset was compiled by hyperspectral imaging of six patients in vivo. All data was collected under a protocol approved by the Institutional Review Board (IRB) of the Johns Hopkins University, School of Medicine. The data was collected using an ASD LabSpec 4 Spectrometer with wavelengths ranging from 450 nm to 1800 nm. During the spectral signature acquisition, the probe was positioned perpendicular to the skin and light was shone in a 10
Conclusions
In this paper we developed a fully-automated and non-invasive method based on a biophysical model of human skin and machine learning. We used this method to estimate the constitutive skin parameters, and in particular accurately measured the thickness of human skin. When comparing the results to Ultrasound sequences obtained from the same anatomical locations and annotated by a physician, we found an error on the order of the gold standard (US); this was further supported by our test of
Conflict of interest statement
None declared.
References (20)
- et al.
Evaluation of the ulnopalmar arterial arches with pulse oximetry and plethysmographycomparison with the Allen׳s test in 1010 patients
Am. Heart J.
(2004) - American Cancer Society, Cancer Facts and Figures,...
- G. Baranoski, A. Krishnaswamy, Light interaction with human skin: from believable images to predictable models, in: ACM...
- C.C. Chang, C.J. Lin, LIBSVM: a library for support vector machines, in: ACM Transactions on Intelligent Systems and...
- et al.
A review of the optical properties of biological tissues
IEEE J. Quantum Electron.
(1990) - E. Claridge, S. Preece, An inverse method for the recovery of tissue parameters from colour images, in: Information...
Research Paper Differentiation of Normal Skin and Melanoma using High Resolution Hyperspectral Imaging
Cancer biology & therapy 5.8
(2006)- et al.
Management of the patient with melanoma
Cancer Skin
(2005) - A. Krishnaswamy, G. Baranoski, A biophysically-based spectral model of light interaction with human skin, in: Computer...
- Krishnaswamy Aravind, Gladimir VG Baranoski, "A study on skin optics." Natural Phenomena Simulation Group, School of...
Cited by (21)
Hyperspectral imaging for non-invasive blood oxygen saturation assessment
2024, Photodiagnosis and Photodynamic TherapyHyperspectral image analysis for rapid and accurate discrimination of bacterial infections: A benchmark study
2017, Computers in Biology and MedicineCitation Excerpt :Thus a complete and balanced review of the various usages of HSI in the biomedical field is beyond the scope of the present work and we refer the interested reader to available surveys [5,26]. What we observed however is that, while there is a high level of heterogeneity in the way multispectral and hyperspectral images are acquired, visualized and processed for biomedical applications, the few studies that use non-trivial classification approaches, are recent and concentrated on cancer detection [27–33]. Therefore there is a need for differentiation and exploration of machine learning solutions for other biomedical applications of HSI data, and our work can be considered a contribution also in this direction.
Kubelka–Munk Model and Stochastic Model Comparison in Skin Physical Parameter Retrieval
2022, Intelligent Systems, Control and Automation: Science and Engineering