A skin condition detection system and method based on optical coherence tomography

Abstract

The present invention relates to a skin state detection system and method based on optical coherence tomography, the system comprises an image acquisition unit, an image preprocessing unit, a skin parameter measurement unit and a deep learning detection unit connected in sequence, the image acquisition unit acquires the original three-dimensional OCT image data of the detected skin area; the image preprocessing unit performs noise reduction and enhancement processing on the original three-dimensional OCT image data to obtain a high-quality three-dimensional OCT image; the skin parameter measurement unit extracts thickness layer parameter features and blood vessel parameter features from the high-quality three-dimensional OCT image; the deep learning detection unit fuses one or more of the thickness layer parameter feature data, the blood vessel parameter feature data, the depth feature of the B-scan tomographic image, and the depth feature of the en-face image to output the state information of the detected skin area. Compared with the prior art, the present invention combines the optical coherence tomography and the deep learning method, and can non-invasively detect the state information of the deep layer of the skin.

Description

Technical Field

[0001] The present invention relates to the technical field of skin condition detection, and in particular to a skin condition detection system and method based on optical coherence tomography. Background Art

[0002] The skin is the largest organ in the human body and the organ that best represents human age. However, due to the influence of internal and external factors such as work and living environment, skin care habits, etc., skin age is often different from actual age. With the development of society, people are paying more and more attention to the health of their skin, and huge market demand has emerged in the fields of dermatology and medical beauty. At present, there is an urgent need for technical solutions that can accurately output information on skin health status and its changes, so that people can easily know their current skin age and other skin quality indicators, and can assist in evaluating the effects of skin care products or cosmetics, so as to flexibly adjust skin care strategies.

[0003] The existing skin health status detection systems on the market, including the American VISIA and Japanese Shiseido IPSA skin testers, mainly focus on analyzing the surface images of the skin and can only obtain information such as skin wrinkles, pores, spots, elasticity, etc., and rarely involve the extraction and processing of deeper information below the basal layer of the skin, including the dermis. At the same time, due to the limitation of resolution, it is difficult and unreliable to accurately measure important skin status information such as skin age and dryness and oiliness through the above skin status detection systems. At the same time, the above skin status detection systems are mainly limited to the detection of facial skin conditions, and cannot detect the skin conditions of other parts of the body, such as hands. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a skin condition detection system and method based on optical coherence tomography, which non-invasively extracts the ultrastructure information of the deep layer of the skin to output the current skin condition information.

[0005] The object of the present invention can be achieved by the following technical solutions: a skin condition detection system based on optical coherence tomography, comprising an image acquisition unit, an image preprocessing unit, a skin parameter measurement unit and a deep learning detection unit connected in sequence, wherein the image acquisition unit is used to obtain original three-dimensional OCT (Optical Coherence Tomography) image data of the detected skin area, wherein the three-dimensional OCT image data comprises a B-scan tomographic image stacked along the y-axis direction and an en-face image stacked along the z-axis direction;

[0006] The image preprocessing unit is used to perform noise reduction and enhancement processing on the original three-dimensional OCT image data to obtain a high-quality three-dimensional OCT image;

[0007] The skin parameter measurement unit is used to extract thickness layer parameter features and blood vessel parameter features from the high-quality three-dimensional OCT image;

[0008] The deep learning detection unit is used to fuse one or more of the thickness layer parameter feature data, the blood vessel parameter feature data, the depth features of the B-scan tomographic image, and the depth features of the en-face image to output the status information of the detected skin area.

[0009] Furthermore, the image acquisition unit includes an FD-OCT imaging device and a storage device.

[0010] Further, the skin parameter measurement unit includes a layering processing module and an OCTA angiography module, wherein the layering processing module is used to layer the high-quality three-dimensional OCT image to distinguish the stratum corneum, epidermis, and dermis structures in the three-dimensional skin data, and determine the positions of the upper boundary of the epidermis, the lower boundary of the stratum corneum, and the dermis-epidermis connection line to obtain one or more thickness layering parameter features including the stratum corneum thickness, epidermis thickness, and dermis thickness parameters of the skin;

[0011] The OCTA angiography module is used to use the OCTA angiography algorithm on continuous B-scan images to obtain angiography en-face images of each skin layer, and obtain one or more vascular parameter characteristics of vascular parameters such as vascular density and vascular perfusion area based on the OCTA images of each cortex.

[0012] Further, the deep learning detection unit includes one or more of a B-scan feature extraction network, a vascular angiography en-face image feature extraction network, a fusion network, and an output network, wherein the B-scan feature extraction network is used to extract B-scan depth feature data from the B-scan tomographic image;

[0013] The angiography en-face image feature extraction network is used to extract depth feature data from each cortical angiography en-face image;

[0014] The fusion network is used to fuse one or more of B-scan depth feature data, angiography en-face depth feature data, thickness layer parameter feature data, and vascular parameter feature data to obtain fused feature data;

[0015] The output network determines and outputs the skin condition information of the detected skin area based on the fused feature data.

[0016] Optionally, the fused network is a fully connected network.

[0017] Furthermore, the skin condition information includes one or more information of skin age, skin texture, collagen content, elastin content, skin roughness, skin moisture content, skin transmittance, skin oil secretion, skin qi and blood, stratum corneum thickness, skin aging degree, and hair follicle health.

[0018] A method for detecting skin condition based on optical coherence tomography comprises the following steps:

[0019] S1, the image acquisition unit obtains the original three-dimensional OCT image data of the detected skin area and outputs it to the image preprocessing unit;

[0020] S2, the image preprocessing unit performs noise reduction and enhancement processing on the original three-dimensional OCT image in sequence to obtain a high-quality three-dimensional OCT image;

[0021] S3, the skin parameter measurement unit segments the high-quality three-dimensional OCT image and determines the skin surface segmentation line to obtain one or more thickness stratification parameter features including skin stratum corneum thickness, epidermis thickness, and dermis thickness parameters;

[0022] According to the skin surface segmentation line, the layered high-quality three-dimensional OCT image is flattened so that the flattened three-dimensional OCT image data has a uniform size;

[0023] Using the OCTA angiography algorithm on the continuous B-scan images to obtain the angiography en-face images of each skin layer, and obtaining one or more vascular parameter features of vascular density, vascular perfusion area vascular parameters according to the en-face images of the blood vessels of each cortical layer;

[0024] S4. Input one or more data of the image, thickness stratification parameter features, and blood vessel parameter features obtained in the above steps into a pre-trained deep learning detection unit to output status information of the detected skin area.

[0025] Further, the thickness layered parameter feature is specifically a local average thickness, a global average thickness or a plurality of matrices containing thickness parameters in all A-lines in the three-dimensional data;

[0026] The blood vessel density in the blood vessel parameter characteristics includes blood vessel area density and blood vessel skeleton density.

[0027] Furthermore, the deep learning detection unit is specifically a convolutional neural network, including but not limited to deep convolutional CNN networks such as Densenet, AlexNet, GoogleNet, VGG-19, VGG-16, ResNet, and optimized network structures;

[0028] The deep learning detection unit includes a deep feature extraction network, a fusion network and an output network connected in sequence, wherein the deep feature extraction network includes one or more of a B-scan feature extraction network and a dermal angiography en-face image feature extraction network;

[0029] The B-scan feature extraction network is used to extract B-scan depth feature data from the B-scan tomographic image; the dermal angiography en-face image feature extraction network is used to extract depth feature data from the dermal angiography en-face image.

[0030] Furthermore, the training process of the deep learning detection unit is specifically as follows: using an image acquisition unit, an image preprocessing unit and a skin parameter measurement unit to obtain a skin data sample set of people with healthy skin;

[0031] Annotating the skin data sample set, wherein the annotation content includes but is not limited to one or more characteristics of skin age, skin texture, collagen content, and elastin content, to obtain a training data set;

[0032] The deep learning detection unit is trained using the training data set, and the network parameters in the deep learning detection unit are optimized based on the difference between the results output by the deep learning detection unit and the skin state information in the training data set.

[0033] Compared with the prior art, the present invention combines optical coherence tomography with deep learning methods to propose a skin detection scheme that non-invasively extracts deep ultrastructure information of the skin and outputs current skin status information. By setting an image acquisition unit, an image preprocessing unit, a skin parameter measurement unit and a deep learning detection unit, the image acquisition unit is used to obtain the original three-dimensional OCT image data of the detected skin area; the image preprocessing unit is used to perform noise reduction and enhancement processing on the original three-dimensional OCT image data to obtain a high-quality three-dimensional OCT image; the skin parameter measurement unit is used to extract thickness stratification parameter features and vascular parameter features from the high-quality three-dimensional OCT image; the deep learning detection unit is used to output the status information of the detected skin area through feature fusion, thereby enabling non-invasive detection of skin status information.

[0034] In the present invention, the skin parameter measurement unit can distinguish the stratum corneum, epidermis, dermis and other structures in the three-dimensional skin data by stratifying the high-quality three-dimensional OCT image; and by using the OCTA angiography algorithm, the angiography en-face image of each layer of the skin is obtained. In this way, the quantitative analysis of the key indicator data of the deep layer of the skin is realized, which is no longer limited to the detection of the surface state of the facial skin, but can also detect the skin state of other parts of the body. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0036] Figure 2 It is a schematic diagram of the method flow of the present invention;

[0037] Figure 3 It is a schematic diagram of the working process of the image preprocessing unit;

[0038] Figure 4 It is a schematic diagram of the working process of the skin parameter measurement unit;

[0039] Figure 5 A schematic diagram of a model training process for performing a noise reduction operation by an image preprocessing unit in an embodiment;

[0040] Figure 6 A schematic diagram of a model testing process in which the image preprocessing unit performs a noise reduction operation in an embodiment;

[0041] Figure 7 A schematic diagram of a model training process in which an image preprocessing unit performs an enhancement operation in an embodiment;

[0042] Figure 8 It is a schematic diagram of the process of layering operation performed by the skin parameter measurement unit in the embodiment;

[0043] Fig. 9 Schematic diagram of the network structure of the deep learning detection unit in the embodiment;

[0044] Description of the marks in the figure: 1. Image acquisition unit, 2. Image preprocessing unit, 3. Skin parameter measurement unit, 4. Deep learning detection unit. DETAILED DESCRIPTION

[0045] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example

[0047] This technical solution combines optical coherence tomography with deep learning methods to achieve non-invasive deep detection of skin conditions. Optical Coherence Tomography (OCT) technology is a non-invasive, high-resolution, and high-sensitivity optical imaging technology. By collecting the backscattered light of the sample tissue and interfering with the reference light, the light scattering rate information in the depth direction of the sample tissue can be obtained in real time. The resolution can reach 1 to 10 microns and the imaging depth can reach 2 to 5 mm. OCT can obtain high-resolution cross-sectional optical images of the skin and has been applied to the study of human skin morphology and in vivo subcutaneous surface structures, such as melanoma imaging and tumor microvascular imaging. It is very suitable for evaluating skin health status, quantifying skin changes, and checking other indications. In addition, in recent years, there has been great progress in the imaging of skin burn tissue, medical cosmetic evaluation, and handheld OCT probe research. OCT provides a fast and effective optical imaging technology for deep skin state detection, and has great potential and broad application prospects in skin care fields such as skin quality detection.

[0048] like Figure 1 As shown, a skin condition detection system based on optical coherence tomography proposed by the present invention comprises an image acquisition unit 1, an image preprocessing unit 2, a skin parameter measurement unit 3 and a deep learning detection unit 4 connected in sequence, wherein the image acquisition unit is used to obtain original three-dimensional OCT image data of the detected skin area, and the three-dimensional OCT image data comprises B-scan tomographic images stacked along the y-axis direction and en-face images stacked along the z-axis direction;

[0049] The image preprocessing unit is used to perform noise reduction and enhancement processing on the original three-dimensional OCT image data to obtain a high-quality three-dimensional OCT image;

[0050] The skin parameter measurement unit is used to extract thickness layer parameter features and vascular parameter features from high-quality three-dimensional OCT images;

[0051] The deep learning detection unit is used to fuse one or more of the thickness layer parameter feature data, the blood vessel parameter feature data, the depth features of the B-scan tomographic image, and the depth features of the en-face image to output the status information of the detected skin area.

[0052] The specific working process of the above system is as follows: Figure 2As shown, in this embodiment, the optical coherence tomography image acquisition unit includes an FD-OCT imaging device and a computer with a storage function, which is used to obtain the original three-dimensional optical coherence tomography image data of the detected skin area, and the obtained original three-dimensional OCT image data has a uniform size and resolution. The acquisition and storage method of the original three-dimensional OCT image data can be a B-scan tomographic image stacked along the y-axis direction or an en-face image stacked along the z-axis direction according to different OCT imaging systems.

[0053] like Figure 3 As shown, the image preprocessing unit mainly performs noise reduction processing on the original three-dimensional OCT image data, and performs enhancement processing on the three-dimensional OCT image data after noise reduction to obtain a high-quality three-dimensional OCT image. The noise reduction processing and enhancement processing can adopt a method of performing two-dimensional processing on each B-scan in the three-dimensional OCT image data frame by frame, or a three-dimensional processing method of performing overall noise reduction and enhancement on the sample data.

[0054] Noise reduction processing can adopt one or more of the following methods: multi-frame averaging method; classic Gaussian filtering, median filtering, mean filtering methods; spatial adaptive filtering methods such as Lee filter and Kuan filter; frequency domain filtering methods based on wavelet transform, Wiener filtering, etc.; diffusion-based filtering methods such as anisotropic diffusion filter; statistical methods based on divergence regularization, Bayesian estimation, etc., and denoising methods based on deep learning, so as to eliminate the multiplicative speckle noise and background noise of the image.

[0055] The enhancement processing can adopt one or more methods such as adaptive contrast enhancement, deconvolution method, sparse reconstruction method based on compressed sensing and deep learning super-resolution reconstruction method based on Generative Adversarial Networks (GAN), so as to improve the contrast or resolution of the image.

[0056] like Figure 4 As shown, the skin parameter measurement unit mainly includes the following steps:

[0057] A) Extract skin layer features:

[0058] The high-quality three-dimensional OCT images are layered to distinguish the stratum corneum, epidermis, dermis and other structures in the three-dimensional skin data. In the layering process, methods including but not limited to edge extraction methods based on gradients and image segmentation methods based on deep learning can be used. In the image segmentation methods based on deep learning, methods including but not limited to U-net, U 2 -net and other deep learning network models.

[0059] Through layering, the positions of the segmentation lines such as the upper boundary of the epidermis, the lower boundary of the stratum corneum, and the dermis-epidermis junction (DEJ) are determined, so that one or more thickness parameters including the thickness of the stratum corneum, the thickness of the epidermis, and the thickness of the dermis can be obtained. The thickness parameter can be the local average thickness, the global average thickness, or a matrix containing the thickness parameters of all A-lines in the three-dimensional data. In order to improve the signal-to-noise ratio of the B-scan image, the high-quality three-dimensional OCT image after layering is flattened according to the skin surface segmentation line, so that the flattened three-dimensional OCT image data has a uniform size.

[0060] B) Extract skin microvascular features:

[0061] The OCTA angiography algorithm is used on the continuous B-scan images to obtain the angiography en-face images of each skin layer. Based on the en-face images of the blood vessels in each cortical layer, one or more of the vascular parameters such as vascular density and vascular perfusion area are obtained. The vascular density includes vascular area density and vascular skeleton density.

[0062] Finally, the deep learning detection unit inputs the feature data or image data obtained in the above steps into the deep learning detection model to obtain the state information of the detected skin area. Among them, the deep learning detection unit includes one or more of a B-scan feature extraction network, a dermal angiography en-face image feature extraction network, a fusion network and an output network. The B-scan feature extraction network is used to extract B-scan feature data from the B-scan tomographic image, the dermal angiography en-face image feature extraction network is used to extract feature data from the dermal angiography en-face image, the fusion network is used to fuse the B-scan feature data, the dermal angiography en face image feature data and at least one of the skin stratification thickness parameter feature and the dermal blood vessel parameter feature to obtain fused feature data, and the output network is used to determine the skin state information of the detected skin area based on the fused feature data. Optionally, the fusion network is a fully connected network.

[0063] In practical applications, it is necessary to pre-train the deep learning detection unit model:

[0064] The skin data of a large number of people with healthy skin are obtained by using an image acquisition unit, an image preprocessing unit and a skin parameter measurement unit;

[0065] According to the actual situation of the collected skin data samples, one or more characteristics including but not limited to skin age (based on the healthy skin condition of the sample population, the default skin age is the actual age), skin texture (dry skin, oily skin, neutral or sensitive skin, etc.), collagen content, elastin content, etc. are respectively marked, wherein the data of collagen content and elastin content can be obtained by calculating the second harmonic generation (SHG) in multi-photon laser scanning imaging technology and the autofluorescence (AF) signal of the superficial dermis;

[0066] The training data is used to train the deep learning detection model of skin OCT images, and the parameters of the machine learning model are optimized according to the difference between the skin status information output by the output network and the skin status information in the training data.

[0067] In addition, the deep learning detection unit can use convolutional neural networks, including but not limited to deep convolutional CNN networks such as Densenet, AlexNet, GoogleNet, VGG-19, VGG-16, and ResNet, and optimized network structures. The deep learning detection unit is used to extract and fuse one or more of the high-quality B-scan images obtained by preprocessing and the skin layer characteristics, skin layer angiography en-face images, skin microvascular characteristics, etc. obtained by the skin parameter measurement unit, and finally the fused data is used as the input of the deep learning model to obtain the final skin state detection result, which contains at least one of the following information: skin age, skin texture (dry skin, oily skin or sensitive skin, etc.), collagen content, elastin content, skin roughness, skin moisture content, skin transmittance, skin oil secretion, skin Qi and blood, stratum corneum thickness, skin aging degree, and hair follicle health.

[0068] This embodiment applies the above technical solution to provide a skin age determination system based on optical coherence tomography and deep learning. The purpose is to collect skin optical coherence tomography data of a specific area of ​​the subject and determine the skin age of the subject in the specific skin area through a deep learning method. This embodiment uses the hand as the specific skin area for detection.

[0069] The technical solution of this embodiment consists of four parts: an optical coherence tomography image acquisition unit, an image preprocessing unit, a skin parameter measurement unit, and a deep learning detection unit.

[0070] The optical coherence tomography image acquisition unit uses a self-developed OCT skin imaging system. The system uses a weak coherence SLED with a central wavelength of 840nm and a bandwidth of 100nm as the illumination light source and a 90:10 broadband fiber coupler as the core of the Michelson interferometer. The sample arm of the system consists of a fiber collimator, a pair of two-dimensional scanning galvanometers, and a near-infrared objective lens with a focal length of 54mm. The optical power of the imaging beam is set to ~800μW, which is much lower than the maximum allowable exposure power specified by the ANSI (American National Standards Institute) standard. The reference arm of the system consists of a fiber collimator, a dispersion compensation unit, a neutral density filter, and a plane mirror mounted on a micro manual translation stage. An additional fiber polarization controller is used to optimize the interference amplitude of the system. The system is controlled by a LabVIEW program and uses a self-developed spectrometer module to realize the acquisition of interference spectra. The spectrometer module uses a 2048-pixel monochrome line scanning CMOS camera (OCTOPLUS CL, e2v) with a maximum readout rate of 250kHz, and cooperates with a high-speed image acquisition card to achieve high-speed OCT imaging. The final system can obtain OCT tomographic imaging with an axial resolution better than 10μm (in air), an imaging depth>1mm, and a field of view greater than 10mm x 10mm. The imaging speed can reach the fastest 500Hz two-dimensional line scanning and 1Hz three-dimensional volume scanning. The three-dimensional original OCT skin image data of the skin area of ​​500 subjects with healthy skin of different ages were obtained through the image acquisition unit as a training data set.

[0071] The image preprocessing unit includes the following steps:

[0072] Step 1: De-noise the B-scan image obtained by the image acquisition unit

[0073] The noise reduction method is to use a one-dimensional mean filter to perform adaptive spatial filtering on the B-scan image in the z-axis direction. The mask size of the mean filter is 4 pixels, so that random noise can be reduced. In order to eliminate speckle noise, the grayscale of the B-scan image is normalized. The specific processing method is to set a grayscale upper bound U b and a grayscale lower bound L b , use the following grayscale transformation formula to reduce speckle noise:

[0074]

[0075] Where XN is the grayscale value after transformation, and XR is the grayscale value before transformation.

[0076] An unsupervised deep learning denoising strategy based on noise-to-noise (Noise2Noise, N2N) is used to denoise the OCT images. The denoising strategy includes a model training step and a model testing step.

[0077] like Figure 5 As shown in , during the denoising model training process, a noisy B-scan image is input into the deep neural network to generate a predicted B-scan image. The predicted B-scan image is compared with another noisy B-scan image collected from the same part to obtain the loss function, and finally the loss value is used to calculate the gradient and update the network parameters. Figure 6 As shown, in the denoising model testing step, the noisy test set B-scan image is input into the optimized deep neural network to obtain the final denoised B-scan image.

[0078] Step 2: Enhance the B-scan image after noise reduction in step 1

[0079] The Adaptive Contrast Enhancement (ACE) algorithm is used to enhance the image contrast. The core principle of the Adaptive Contrast Enhancement algorithm is the Unsharp Mask technique, which refers to obtaining the low-frequency Unsharp Mask part of the image through a low-pass filter, and then obtaining the high-frequency part by subtracting the original image from the low-frequency part point by point. Since the details of the image are often contained in the high-frequency component, the composite image obtained by multiplying the high-frequency component by the contrast gain and adding the low-frequency mask component is the result of image enhancement.

[0080] f(i,j)=m l (i,j)+CG·m h (i,j)

[0081] In the formula, m h (i,j) represents the high frequency component, m l (i, j) represents the low-frequency component, and CG represents the contrast gain.

[0082] The difference between the adaptive contrast enhancement algorithm and the ordinary linear unsharp masking technique lies in the selection of the contrast gain CG. The linear unsharp masking technique often uses a fixed constant as the contrast gain coefficient. The value of CG in the ACE algorithm is given by the following formula:

[0083]

[0084] Where D is the global standard deviation, σ x (i,j) is the local standard deviation of the original image in the area centered at x(i,j) and with a size of (2n+1)× (2n+1).

[0085] This embodiment uses a deep learning method based on GAN or a dictionary learning method based on sparse representation to perform super-resolution reconstruction on OCT images, thereby enhancing the resolution of skin OCT images. The training process of the deep learning super-resolution reconstruction model based on GAN is as follows: Figure 7 As shown, the high-resolution OCT image uses sample data with the maximum sampling rate that can be collected by the OCT image acquisition unit, and the low-resolution OCT image used in the training set is obtained by downsampling the above high-resolution OCT image.

[0086] The skin parameter measurement unit includes the following steps:

[0087] Step 1: Extract skin stratification features:

[0088] In this embodiment, the skin stratification process of the parameter feature extraction step is as follows: Figure 8 shown.

[0089] The convolutional neural network is used to segment the high-quality B-scan images that have been denoised and enhanced to obtain the location of the skin surface and DEJ, and the thickness of the epidermis of the skin sample. The segmentation method includes three steps: data screening, model training, and model optimization. Data screening is to randomly select all B-scan images of 20 sample body data to manually mark the skin surface segmentation line and the dermis-epidermis segmentation line, and the remaining samples are used as test set data;

[0090] In this embodiment, the model training step adopts U-net network for training. U-net is a fully connected network (FCN) that is widely used in image segmentation. Compared with other neural networks, FCN network only contains convolution layer and pooling layer. Compared with the fully connected layer, the convolution layer only reflects the information inside the convolution kernel, so it has fewer parameters. In order to better utilize the three-dimensional OCT data to improve the training efficiency, a multi-directional training method is used. This method increases the amount of data in the training set by measuring the cross-sectional images of the three-dimensional OCT along the x-axis and y-axis directions respectively.

[0091] The model optimization step uses the adaptive momentum estimation (Adam) optimizer, which is a gradient-based optimizer with a random objective function. The optimized U-net model is used to predict the skin surface segmentation line and dermis-epidermis connection line of the B-scan test set.

[0092] For each test data set, the probability of each pixel belonging to skin tissue is first predicted along the y direction, and then the process is repeated along the x direction. Through the multi-directional U-net network, two output volume data can be obtained, and then the B-scan in these volume data is converted into a non-directional graph, and the weight is determined by the two volume data respectively. Each pixel in the volume data is regarded as an eight-connected node, and then the Dijkstra shortest path algorithm is used to obtain the skin surface segmentation line and the dermis-epidermis connection line. Finally, according to the stratification of the entire three-dimensional OCT sample image, the skin stratification parameters such as the average epidermal thickness and stratum corneum thickness of the sample are obtained.

[0093] Step 2: Flatten the layered high-quality three-dimensional OCT image according to the skin surface segmentation line, and make the flattened three-dimensional OCT image data have a uniform size. The specific stretching method is:

[0094] The skin surface segmentation line is represented as a series of vertical coordinates Y = [Y1, Y2, ..., Yn], where n represents the width of the B-scan. The normalized image is obtained by the following formula:

[0095] Im flatten (x,y)=Im original (x,yY x )

[0096] In the formula, Im flatten (x, y) represents the stretched pixel, Im original (x,y) represents the original image pixel, Y x Indicates the vertical coordinate of the split line in column x.

[0097] Step 3: Extract skin microvascular features:

[0098] The OCTA angiography algorithm is used on the continuous B-scan images to obtain the en-face image of the angiography in the dermis. Based on the en-face image of the dermis blood vessels obtained by the OCTA angiography algorithm, one or more of the vascular parameters such as the vascular density and the vascular perfusion area are obtained. The vascular density includes the vascular area density and the vascular skeleton density.

[0099] The OCTA angiography algorithm obtains B-scan images of the same location in the skin area through continuous scanning, and obtains decorrelation information in the skin cross-section at that location through the separated spectral amplitude decorrelation method (SSADA), thereby obtaining OCTA three-dimensional vascular data with low signal-to-noise ratio.

[0100] The blood vessel parameters can be measured by performing edge detection on the obtained dermal angiography en-face image, thereby obtaining the number of blood vessels per unit area.

[0101] The deep learning detection unit uses at least one of the OCT skin image or feature data obtained by the above operation as input, and estimates the skin age of the detected skin area through a deep learning detection model.

[0102] Among them, the training method of the deep learning detection unit is as follows:

[0103] The first step is to annotate the age of 500 healthy OCT skin images collected according to the actual situation of the samples. 400 samples are randomly selected from the annotated OCT 3D data samples as the training data set, and the other 100 data samples are used as the test data set. The 400 training data samples are required to cover all samples from 10 to 80 years old and the number of samples for each age must be balanced. In addition, data enhancement can be performed by changing the contrast of OCT images to expand the training data set.

[0104] In the second step, the high-quality B-scan images, skin epidermal thickness information, dermal OCTA angiography en-face images, and vascular information obtained from the training data set through the previous steps were used as input to train the skin OCT image deep learning detection model. The parameters of the machine learning model were optimized based on the difference between the skin age output by the output network and the skin age in the training data.

[0105] In an embodiment of the present invention, a deep learning detection unit is used to perform feature fusion on one or more of the preprocessed high-quality B-scan images, epidermal thickness parameters obtained through parameter feature extraction, dermal angiography en-face images, and dermal vascular parameters, and finally the fused feature data is used as the input of the deep learning model to obtain the final skin disease detection results.

[0106] The deep learning detection unit can adopt a convolutional neural network, including but not limited to Densenet, AlexNet, GoogleNet, VGG-19, VGG-16, ResNet and other deep convolutional CNN networks and optimized network structures.

[0107] The deep learning detection unit includes one or more of a B-scan feature extraction network, a dermal angiography en-face image feature extraction network, a fusion network, and an output network.

[0108] The B-scan extraction network is used to extract B-scan feature data from the B-scan tomographic image, the dermal angiography en-face image skin parameter measurement network is used to extract feature data from the dermal angiography en-face image, the fusion network is used to fuse the B-scan feature data, the dermal angiography en-face image feature data and at least one of the skin layer thickness parameter feature and the dermal layer blood vessel parameter feature to obtain fused feature data, and the output network is used to determine the skin age of the detected skin area based on the fused feature data. The network structure of the deep learning detection unit used in the embodiment of the present invention is as follows: Fig. 9 shown.

[0109] In summary, this technical solution combines optical coherence tomography with deep learning methods to provide a skin detection solution that non-invasively extracts deep ultrastructural information of the skin, thereby outputting current skin condition information. It allows users to accurately grasp their current skin condition and evaluate the impact of cosmetics or skin care products on skin condition, so as to flexibly adjust skin care methods and improve skin health. Optical coherence tomography skin imaging technology can obtain skin condition information non-invasively; artificial intelligence skin image processing and analysis technology can enhance image quality, extract skin structure and tissue composition information, obtain vascular tomography images, and realize quantitative analysis of key indicators; skin condition clinical data is used to achieve skin condition detection technology with high sensitivity and specificity, which can more conveniently, quickly and accurately detect various skin condition parameters of the tested skin area of ​​the subject, and can provide a reference for consumers to adjust their skin care methods and measure the effectiveness of skin care products.

Claims

1. A skin condition detection system based on optical coherence tomography, characterized in that: The device comprises an image acquisition unit (1), an image preprocessing unit (2), a skin parameter measurement unit (3) and a deep learning detection unit (4) which are connected in sequence, wherein the image acquisition unit (1) is used to obtain original three-dimensional OCT image data of the detected skin area, wherein the three-dimensional OCT image data comprises B-scan tomographic images stacked along the y-axis direction and en-face images stacked along the z-axis direction; The image preprocessing unit (2) is used to perform noise reduction and enhancement processing on the original three-dimensional OCT image data to obtain a high-quality three-dimensional OCT image; The skin parameter measurement unit (3) is used to extract thickness layer parameter features and blood vessel parameter features from the high-quality three-dimensional OCT image; The deep learning detection unit (4) is used to fuse the thickness layer parameter feature data, the blood vessel parameter feature data, the depth feature of the B-scan tomographic image, and the depth feature of the en-face image to output the state information of the detected skin area; The skin parameter measurement unit (3) comprises a layer processing module and an OCTA angiography module, wherein the layer processing module is used to layer the high-quality three-dimensional OCT image to distinguish the stratum corneum, epidermis, and dermis structures in the three-dimensional skin data, determine the positions of the upper boundary of the epidermis, the lower boundary of the stratum corneum, and the dermis-epidermis connection line, and obtain thickness layer parameter characteristics including stratum corneum thickness, epidermis thickness, and dermis thickness parameters; The OCTA angiography module is used to use the OCTA angiography algorithm on the continuous B-scan images to obtain the angiography en-face images of each layer of the skin, and obtain the vascular density and vascular perfusion area vascular parameters according to the OCTA images of each cortex; The deep learning detection unit (4) comprises a B-scan feature extraction network, a vascular angiography en-face image feature extraction network, a fusion network and an output network, wherein the B-scan feature extraction network is used to extract B-scan depth feature data from the B-scan tomographic image; The angiography en-face image feature extraction network is used to extract depth feature data from each cortical angiography en-face image; The fusion network is used to fuse B-scan depth feature data, angiography en-face depth feature data, thickness layer parameter feature data, and blood vessel parameter feature data to obtain fused feature data; The output network determines and outputs the skin condition information of the detected skin area based on the fused feature data.

2. The skin condition detection system based on optical coherence tomography according to claim 1, characterized in that: The image acquisition unit (1) comprises an FD-OCT imaging device and a storage device.

3. The skin condition detection system based on optical coherence tomography according to claim 1, characterized in that: The fused network is a fully connected network.

4. The skin condition detection system based on optical coherence tomography according to claim 1, characterized in that: The skin condition information includes one or more information of skin age, skin texture, collagen content, elastin content, skin roughness, skin moisture content, skin light transmittance, skin oil secretion, skin Qi and blood, stratum corneum thickness, skin aging degree, and hair follicle health.

5. A skin condition detection method using the skin condition detection system based on optical coherence tomography as claimed in claim 1, characterized in that: The following steps are involved: S1, the image acquisition unit obtains the original three-dimensional OCT image data of the detected skin area and outputs it to the image preprocessing unit; S2, the image preprocessing unit performs noise reduction and enhancement processing on the original three-dimensional OCT image in sequence to obtain a high-quality three-dimensional OCT image; S3, the skin parameter measurement unit segments the high-quality three-dimensional OCT image and determines the skin surface segmentation line to obtain thickness stratification parameter characteristics including skin stratum corneum thickness, epidermis thickness, and dermis thickness parameters; According to the skin surface segmentation line, the layered high-quality three-dimensional OCT image is flattened so that the flattened three-dimensional OCT image data has a uniform size; The OCTA angiography algorithm is used on the continuous B-scan images to obtain the angiography en-face images of each skin layer, and the vascular density and vascular perfusion area vascular parameters are obtained according to the en-face images of the blood vessels in each cortical layer as vascular parameter features; S4. Input the image, thickness layer parameter features, and blood vessel parameter features obtained in the above steps into a pre-trained deep learning detection unit to output status information of the detected skin area.

6. A skin condition detection method according to claim 5, characterized in that: The thickness layered parameter feature is specifically a local average thickness, a global average thickness or a plurality of matrices containing thickness parameters in all A-lines in the three-dimensional data; The blood vessel density in the blood vessel parameter feature includes blood vessel area density and blood vessel skeleton density.

7. A skin condition detection method according to claim 6, characterized in that: The deep learning detection unit is specifically a convolutional neural network, including Densenet, AlexNet, GoogleNet, VGG-19, VGG-16, ResNet deep convolutional CNN network and optimized network structure; The deep learning detection unit includes a deep feature extraction network, a fusion network and an output network connected in sequence, and the deep feature extraction network includes a B-scan feature extraction network and a dermal angiography en-face image feature extraction network; The B-scan feature extraction network is used to extract B-scan depth feature data from the B-scan tomographic image; the dermal angiography en-face image feature extraction network is used to extract depth feature data from the dermal angiography en-face image.

8. A skin condition detection method according to claim 7, characterized in that: The training process of the deep learning detection unit is specifically as follows: using an image acquisition unit, an image preprocessing unit and a feature extraction unit to obtain a skin data sample set of people with healthy skin; Label the skin data sample set, including skin age, skin texture, collagen content, and elastin content, to obtain a training data set; The deep learning detection unit is trained using the training data set, and the network parameters in the deep learning detection unit are optimized based on the difference between the results output by the deep learning detection unit and the skin state information in the training data set.

Images