Method and system for the segmentation of a cross-sectional skin image

Abstract

A computer-implemented method for the segmentation of a cross-sectional skin image, comprising: - obtaining (12, 50) the cross-sectional skin image; - processing (20, 54) the cross-sectional skin image, said processing comprising edge detection on the image and / or contrast enhancement of the image; segmenting (22, 60) the processed image by an unsupervised classifier to identify one or more skin layers.

Description

METHOD AND SYSTEM FOR SEGMENTING A CROSS-SECTIONAL SKIN IMAGE DESCRIPTION technical field

[0001] This presentation concerns the field of computer image processing applied to dermatological analysis, and more specifically a system and a method for segmenting a cross-sectional image of skin. Previous technique

[0002] Currently available imaging techniques allow for the non-invasive acquisition of cross-sectional images of an individual's (human or animal) skin, that is, a cross-sectional view of the skin from its surface in contact with the external environment down to the deeper layers. One example of such techniques is LC-OCT, short for Line-field confocal optical coherence tomography, which can produce an image in two dimensions (2D) or three dimensions (3D).

[0003] Various applications require the analysis of the resulting cross-sectional images, for example, determining skin age, evaluating the effectiveness of a skin treatment, etc. Currently, such analysis is performed visually by experts, which is time-consuming, expensive, and introduces human error.

[0004] Given the advancements in computer vision techniques, training a machine learning model to perform this analysis was considered. However, training the model relies on a database of manually annotated skin images. Thus, in addition to suffering from the aforementioned drawbacks, such a database is time-consuming to compile, and the model's generalization capabilities are limited, for example, to the ethnic groups included in the training data.

[0005] Therefore, there is a need for a new type of process and system for segmenting a cross-sectional image of skin. Description of the invention

[0006] To this end, the present presentation concerns a computer-based method for segmenting a cross-sectional skin image, comprising: - obtaining the transverse skin image; - the processing of the transverse skin image, said processing including the detection of contours on the image and / or the increase of the image contrast; - the segmentation of the processed image by an unsupervised classifier to identify one or more skin layers.

[0007] A cross-sectional skin image can be obtained using any desired technique, preferably non-invasive, such as the previously mentioned LC-OCT. The cross-sectional skin image can be two-dimensional and can optionally be derived from a three-dimensional image. Acquisition may involve acquiring the image from an individual using a dedicated sensor, such as a tomograph, or retrieving a previously acquired image stored in a database, for example, on a local or remote server.

[0008] Edge detection is a process that takes a cross-sectional image of skin as input and returns, as output, information representing the edges in the input image; this can be the transformed input image in which the edges have been highlighted, as opposed to other image elements, or any information representing the position of the edges in the input image. Edge detection is a well-known processing method.

[0009] Image contrast enhancement is a process that takes as input a cross-sectional skin image and outputs a transformed image in which the color contrasts between the pixels of the input image are accentuated. For the purposes of this discussion, the term "color" of a pixel refers both to colors in the usual sense (red, green, blue, and their compounds) and to shades of gray on a black and white scale. Increasing the contrast of an image is a known processing technique.

[0010] When the processing of a cross-sectional skin image includes both edge detection and contrast enhancement, one of these steps may take the output of the other step as input instead of the cross-sectional image. The cross-sectional image is thus taken as input indirectly. The cross-sectional image is also taken as input indirectly when the cross-sectional skin image undergoes preprocessing. Alternatively, the processing may take the cross-sectional skin image as input directly.

[0011] Image segmentation involves locating different semantic instances within the image, each corresponding to a specific object or area to be detected. This operation is well-known and can be implemented by a machine learning model, such as a classifier. The classifier takes the processed image as input and identifies coherent areas within it, classifying them into categories defined during the classifier's training. Classifiers for semantic image segmentation are well-established models, sometimes based on one or more artificial neural networks, such as a convolutional neural network (CNN), or on one or more random forests.For the purposes of this presentation, the terms learning and training are used interchangeably to refer to the process of calibrating the classifier's internal parameters (weights and biases, typically) to examples initially given before using the classifier in the so-called inference phase.

[0012] In this case, the classifier is unsupervised, meaning it was trained using data that was not manually annotated. This makes preparing the training database faster and avoids annotation bias. The lack of supervised training is compensated for by the fact that the classifier is explicitly guided in its segmentation by the prior image processing. As a result, the segmentation process is more efficient and robust than previous methods.

[0013] In some embodiments, the segmentation process further includes image preprocessing comprising at least one of a normalization, denoising and averaging of several transverse skin images.

[0014] Image preprocessing can occur before the aforementioned processing. Preprocessing aims to avoid artifacts and color variations that could hinder processing or segmentation, even though they are not semantically relevant. For example, normalization aims to bring pixel colors into a predetermined scale, so that the overall image coloring is consistent with the coloring typically processed by the classifier. Normalization thus reduces variations between different images seen by the classifier. Denoising aims to remove noise from the image that may be due to the image acquisition technique. Denoising allows for a numerical increase in image quality, which improves subsequent processing and segmentation.Finally, averaging several cross-sectional skin images, ideally adjacent to one another, eliminates local effects by considering a larger skin area than a single cross-section. This results in a more regular averaged cross-sectional image, which improves subsequent processing and segmentation and increases the representativeness of the segmentation. Averaging also reduces the number of images to be segmented, thus accelerating the process. Normalization, denoising, and averaging can be performed once or several times for each image and combined in any desired order.

[0015] In some embodiments, one or more segmented skin layers are removed from the image, and the processing and segmentation steps are repeated. Thus, it is possible to proceed iteratively by segmenting at least one layer, removing it from the image, and then processing and segmenting the remaining image, and so on as many times as desired. This iterative method is advantageous because some skin layers may be easier to identify than others, whether for imaging reasons (the layers are better defined) or physiological reasons (the location of the layers is better known). The iterative approach allows for more precise identification of even the most difficult-to-segment layers, since the previously segmented layers, which are easier to identify, were removed from the image in a previous iteration.

[0016] Optionally, one or more skin layers are segmented from the more superficial to the deeper layers. The more superficial layers are generally easier to segment due to their interface with the external environment; thus, at least one boundary between these layers can be obtained relatively easily. Overall, the segmentation process is therefore more efficient.

[0017] Optionally, the preprocessing can also be repeated or not during each iteration, with the same preprocessing operation(s) (normalization, averaging, denoising) or one or more different operations, and in the same order or in a different order.

[0018] In some embodiments, the segmentation includes the detection of the location of an interface between two layers by classifying longitudinal skin zones comprised between one or more first longitudinal zones associated with a first of the two layers and one or more second longitudinal zones associated with a second of the two layers, and the segmentation of the interface on the basis of the detected location.

[0019] The longitudinal direction is transverse to the transverse direction and extends generally along the skin. Thus, provided the layers are sufficiently regular, a longitudinal area of ​​skin normally belongs to a single layer.

[0020] In these embodiments, segmentation is based on the fact that one or more initial longitudinal zones can be associated with a first layer with a good degree of certainty, while one or more secondary longitudinal zones, distant from said initial zones, can be associated with a second layer, adjacent to the first layer, with a good degree of certainty. Thus, even if the interface between the first and second layers is difficult to detect a priori, its location can be approximated by associating zones on either side of this interface with the first and second layers. Step by step, the possible location of the interface is thus reduced, which helps guide the segmentation of the interface.The proposed segmentation process thus makes it possible to segment interfaces even if they are a priori difficult to detect, and consequently to segment with precision the skin layers on either side of this interface, namely the first layer and the second layer.

[0021] In some embodiments, the processing includes removing areas of the image that show hair. Indeed, hair can interfere with processing (particularly normalization, as it creates very dark areas) and / or segmentation (especially because it introduces contours that are not layer contours), so it is advantageous to remove areas that show hair. Such areas may include actual hair or show the shadow of hair on the skin image.

[0022] Removing areas of hair from the image can be done alone or followed by reconstructing the affected areas to ensure a hair-free continuity. Reconstruction can be performed by interpolation, particularly polynomial interpolation. However, it is not necessary because, since hair is generally transverse to the skin layers, removing a hairy area usually leaves other visible areas on either side, allowing for the segmentation of the skin layers.

[0023] In some embodiments, these hair-bearing areas are detected on one or more longitudinal skin images. The longitudinal skin image can be two-dimensional and can optionally be derived from a three-dimensional image, typically the same three-dimensional image from which the cross-sectional image is obtained. Detecting hair-bearing areas on a longitudinal image allows for better differentiation of hair from other skin features. The location of a hair-bearing area can be transferred to the original 3D image, if available, and then from the 3D image to the cross-sectional image.

[0024] To facilitate detection, the longitudinal image can be obtained by principal component analysis of several longitudinal images of superficial layers of the skin.

[0025] In some embodiments, these areas showing hairs are detected by a classifier and / or a thresholding mechanism. A classifier, for example based on machine learning, makes it possible to detect a certain continuity in the hairs even in the presence of breaks in the image. For example, a possible classifier is a k-means clustering model, which is well-known per se. A thresholding mechanism, which includes, for example, identifying as hair all pixels with a color exceeding a certain threshold, is, on the other hand, more robust at detecting small hair fragments but does not incorporate a notion of of continuity. Thresholding can be triangular or another type of thresholding, for example, using the Otsu method or the Li method. These methods are well-known. Combining a classifier and a thresholding mechanism allows us to leverage the complementary advantages of these two approaches to improve hair detection.

[0026] In some embodiments, the cross-sectional skin image is obtained from three-dimensional skin imaging data.

[0027] In some embodiments, one or more segmented layers include the stratum corneum (also known by its Latin name stratum corneum) and optionally the living epidermis located between the stratum corneum and the dermo-epidermal junction.

[0028] This presentation also focuses on a system for segmenting a cross-sectional image of skin, configured to: - obtain a cross-sectional image of skin; - process the cross-sectional skin image, said processing including edge detection on the image and / or increasing image contrast; - segment the image processed by an unsupervised classifier to identify one or more skin layers.

[0029] The system may include all or part of the characteristics described above with regard to the process, mutatis mutandis. The system may have the hardware architecture of a computer, and in particular include one or more processors configured to execute the steps above.

[0030] This presentation also concerns a program set containing instructions for executing the steps of the processing procedure as described above when said program set is executed by at least one computer or microprocessor. This program set can use any programming language and be in the form of source code, object code, or code intermediate between source and object code, such as in a partially compiled form, or in any other desirable form.

[0031] This presentation also refers to a computer-readable recording medium on which is recorded at least one computer program including instructions for carrying out the steps of the processing procedure as described above.

[0032] The information storage medium can be any entity or device capable of storing the program. For example, the medium may include a mass storage device, such as a hard drive. More generally, mass storage devices suitable for the tangible incorporation of computer program instructions and data include all forms of non-volatile memory, including, for example, semiconductor memory devices such as ROM, CD-ROM, EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard drives and removable disks; and magneto-optical disks.

[0033] On the other hand, the information medium can be a transmissible medium such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means. The program described herein can, in particular, be downloaded from a network such as the Internet. Brief description of the drawings

[0034] Other features and advantages of the object of this presentation will emerge from the following description of embodiments, given by way of non-limiting examples, with reference to the attached figures.

[0035] Figure 1 is a block diagram illustrating a first step of a segmentation process according to one embodiment.

[0036] Figure 2 is a block diagram illustrating a second step of a segmentation process according to one embodiment.

[0037] Figure 3 is a block diagram illustrating a segmentation system according to one embodiment. Detailed description

[0038] A method for segmenting a transverse skin image according to one embodiment is described with reference to Figures 1 and 2. As previously stated, the segmentation method 10 first comprises obtaining 12 a transverse skin image. The two-dimensional transverse skin image can be obtained from three-dimensional skin imaging data, for example, from a three-dimensional image itself. obtained by a 3D imaging technique such as LC-OCT 3D. As illustrated in Figure 1, LC-OCT 3D returns a 3D image 12a representing a certain volume of skin, and it is possible to extract from this 3D image both transverse skin images (vertical sections on image 12a as illustrated in Figure 1) and longitudinal skin images (horizontal sections on image 12a as illustrated in Figure 1).However, other 3D imaging techniques can be used, for example: interferometric imaging (notably Optical Coherence Tomography), fluorescence imaging (notably laser scanning microscopy such as Confocal Reflectance Microscopy, Confocal Fluorescence Microscopy, and Two- or Multi-Photon Microscopy), or optoacoustic imaging (for example, raster-scanning optoacoustic mesoscopy (RSOM) or multispectral optoacoustic tomography (MSO)). Furthermore, it is also possible to obtain a 2D image directly, without using 3D imaging data.

[0039] In this embodiment, the segmentation process 10 is iterative: the skin layers are segmented one after another, with the layers segmented in a previous iteration being removed from the cross-sectional image for the current iteration. As will be seen later, each iteration can include at least one processing step and at least one segmentation step. This process can be carried out by starting with the more superficial layers and progressively moving towards the deeper layers of the skin, the successively segmented layers preferably being adjacent to one another. Thus, Figure 1 illustrates the segmentation of the outermost layer of the skin, namely the stratum corneum.

[0040] Before the processing described below, the segmentation process 10 may include preprocessing of the cross-sectional skin image. This preprocessing may include at least one normalization, denoising, and averaging of several cross-sectional skin images. In this example, the preprocessing includes all three steps.

[0041] For example, a normalization step 14 involves dividing the value of each pixel by a common value to match the color scale to a predetermined scale. This operation is known as such in theory; in the present context, it allows us to reduce the differences in values ​​between pixels and the variances on certain metrics that can be calculated based on the cross-sectional image of skin once segmented.

[0042] A denoising step 16, which follows normalization 14 but could precede it, removes the noise introduced by the transverse skin image acquisition technique. Denoising 16 can be performed using a filter, for example, a median filter. The median filter has the advantage of preserving contours and curvature in the image, which, as we will see later, is beneficial for subsequent processing. However, other types of filters that preserve image contours can be used, for example, a bilateral filter, an anisotropic diffusion-based filter, a Kuwahara filter, or patch denoising (non-local means). The filter can be chosen by a person skilled in the art, based in particular on the balance between the desired performance and the available computing power.

[0043] Although it is possible to apply the segmentation process 10 to a single cross-sectional skin image, the present embodiment takes advantage of the availability of 3D imaging data to consider and average multiple cross-sectional skin images, thereby increasing the regularity and representativeness of the image. Thus, during an averaging operation 18, a cross-sectional skin image can be reconstructed as the average of at least two, three, four, or five cross-sectional images, ideally consecutive. For the sake of brevity, in this discussion, we refer to a cross-sectional skin image in the singular, even though it may result from a plurality of cross-sectional skin images.

[0044] The cross-sectional skin image undergoes processing 20, which in this case includes edge detection. For this purpose, the segmentation process can employ a detection algorithm based on the Hessian matrix of the image. Such detection is known as such, and the resulting image is illustrated under reference 20a. Edge detection algorithms based on the Hessian matrix make it possible to highlight the contrasting separation between the stratum corneum and the living epidermis, as can be seen in the processed image 20a. However, other edge detection methods could be used, for example, a Sobel filter, a Prewitt filter, a Canny filter, or other filters based on the first derivative, or preferably filters based on the second derivative, such as a Laplacian filter or a Hessian filter, etc. The Hessian matrix-based detection algorithm proves more refined in use and offers the advantage of detecting not only edges but also curves. Overall, the preprocessing limits the disturbances encountered by the edge detection algorithm.

[0045] Once processed, the image is fed into an unsupervised trained classifier to identify one or more skin layers, initially the stratum corneum. The segmentation process thus involves a segmentation 22 in the strict sense, which can employ k-means partitioning (sometimes called k-barycenters) and / or a decision tree forest. These algorithms, known as such and used here in combination, reliably separate the two main parts located on either side of the contour highlighted by the processing 20 in this context. K-means partitioning is an unsupervised classification algorithm that groups together pixels referring to the same skin layer, based on the image 20a (more precisely, on the image of the eigenvalues ​​of the Hessian matrix, if applicable).The number of layers can be specified by the user, knowing the depth of the transverse skin image: typically, in the present embodiment, three layers (therefore three partitioning classes) correspond to the stratum corneum, the living epidermis and the dermis respectively.

[0046] Although k-means partitioning can be used on its own, its results can be improved with a decision tree forest. The forest takes as input the image annotated by k-means partitioning and provides a more complete segmentation of that image.

[0047] The segmentation process results, at step 24, in an image 24a in which the contour of at least one layer, here the stratum corneum, is identified. If necessary, the image 24a can be enhanced by image processing, for example based on mathematical morphology operators, typically non-linear operators such as erosion, dilation, opening, or closing.

[0048] The portion above the contour highlighted by processing step 20 and segmented in segmentation step 22 is the outermost layer of the skin, namely the stratum corneum, while the other portion is formed from the deeper layers of the skin. Thus, a first iteration of the segmentation process allows us to segment, in the cross-sectional image of the skin, the stratum corneum. At this stage, the segmentation of deeper layers returned by the k-means partitioning can be taken as is or can be ignored if it is not considered reliable enough, for example due to the treatment aimed at improving the detectability of the stratum corneum.

[0049] Concurrently, to avoid disruptions in this segmentation, the processing may include removing areas of the image showing hair. In a cross-sectional image, hair is only seen as such if it extends precisely into the plane of the slice, which is rare; most of the time, a cross-sectional image contains only scattered fragments that interfere with segmentation. To facilitate the search for hair, areas showing hair are therefore instead detected in one or more longitudinal skin images. Longitudinal skin images can be obtained from the previously mentioned 3D image 12a.

[0050] To better observe the hairs, a fusion 30 of several longitudinal images can be performed, in particular of several longitudinal images on the side of the skin surface. Such a fusion can include a principal component analysis (PCA), known as such, which performs dimensionality reduction in order to show all the hairs on a single resulting image 30a. However, other techniques can be employed, such as linear discriminant analysis (LDA), singular value decomposition (SVD), non-negative matrix factorization (NMF), independent component analysis (ICA), or autoencoders whose latent space would be examined.

[0051] Based on the resulting image 30a, areas showing hair can be detected, for example, by a classifier and / or a thresholding mechanism. In this case, Figure 1 illustrates step 32 using a k-means partitioning classifier, and step 34 using a thresholding algorithm such as triangular thresholding. As mentioned previously, other classifiers and thresholding methods could be used instead. These two steps are performed independently, and their results, illustrated by images 32a and 34a, are then combined in step 36 to retain only the areas identified as hair. detected as such consistently in both images 32a and 34a. The thresholding mechanism returns fewer errors than the classifier, but the classifier represents hair continuity better than the thresholding mechanism; therefore, their combined use is advantageous. The image resulting from this combination is referenced as 36a.

[0052] The location of the hairs, now known in the longitudinal plane, is transferred to the transverse images using the 3D coordinates of image 12a. This ultimately allows, in step 38, the removal of the P-zones showing hairs from the transverse image. Note that the removed P-zones can extend across the entire height of the transverse image.

[0053] The previously obtained segmentation and the cross-sectional image devoid of hair-bearing areas can be combined in step 40, resulting in image 40a showing three sub-images (more generally, a plurality, namely N+1 where N is the number of non-contiguous hair-bearing areas removed) in which the stratum corneum (SC) is segmented. In image 40a, the contours of the stratum corneum are highlighted. Figure 1 illustrates the removal of hair-bearing areas after segmentation of the stratum corneum; however, it is preferable to perform this removal before assessing the thickness of the stratum corneum, as hairs lead to a slight overestimation of this thickness.Alternatively, areas showing hair can be removed from the image before segmentation, for example from image 12a, or before or after preprocessing or before or after processing, in which case segmentation 22 can be applied to the resulting sub-images.

[0054] Optionally, the areas of hair removed from image 38a or image 40a can be replaced with a hairless skin reconstruction. In particular, the contours can be made continuous using a polynomial reconstruction, for example, an improved asymmetry least squares (ALS) reconstruction.

[0055] Figure 2 illustrates the segmentation of a layer deeper in the skin than the stratum corneum, for example the living epidermis located between the stratum corneum and the dermo-epidermal junction.

[0056] Before the processing described below, the segmentation procedure may also include, at this stage, preprocessing of the cross-sectional skin image. As before, the preprocessing may This involves at least one normalization, denoising, and averaging of several cross-sectional skin images. In this example, the preprocessing includes these three steps, however in a different order than previously shown.

[0057] For example, at step 50, cross-sectional skin images are extracted from the 3D image 12a, optionally normalized, and these cross-sectional images, ideally consecutive, are averaged to reconstruct a more regular averaged cross-sectional image 50a. All or part of the indications given regarding averaging 18 may apply, mutatis mutandis, although a different averaging method may also be used.

[0058] The averaged transverse image 50a can then undergo denoising 52, similar to denoising 16 and also using a median filter. All or part of the instructions given for denoising 16 may apply, mutatis mutandis. However, a different denoising method could be used. Although the preprocessing has been described again for the sake of completeness and to show the independence of the segmentation of the different layers, it is not necessary to repeat the steps that would have already been implemented for the segmentation of the corneal layer, since it is possible to retrieve the corresponding preprocessed images.

[0059] As a processing step 54, the cross-sectional image then undergoes contrast enhancement, for example, by means of contrast-limited adaptive histogram equalization (CLAHE). This technique, known as such, allows in this context for improving contrast in the image while limiting the introduction of noise, which is all the more desirable as this step occurs after denoising 52. The contrast enhancement then serves as a guide for the subsequent unsupervised segmentation.

[0060] However, before segmenting the living epidermis, the previously segmented layers, in this case the stratum corneum, are removed from the image. This removal can be performed by simply subtracting the segmented stratum corneum from the transverse image, or more simply, by trunculating the transverse image just below the lowest point of the stratum corneum. Figure 2 thus illustrates that the processed image 56a, unlike the previous images, no longer includes the stratum corneum, the hair follicles, and the free space above the stratum corneum.

[0061] The processed image 56a can be segmented by a supervised or unsupervised classifier to identify the dermo-epidermal junction and, consequently, the living epidermis located between the stratum corneum and the dermo-epidermal junction. In one example, this segmentation begins with a location detection step 58. In this step, longitudinal skin zones are classified between a first longitudinal zone Z1, clearly associated with the living epidermis (because it is located just below the stratum corneum, which has been identified and optionally removed), and a second longitudinal zone Z2, clearly associated with the dermis (because it corresponds to the deepest layers of the cross-sectional image).More specifically, a forest of decision trees (or any other classifier, for example a support vector machine, although the latter is more computationally expensive) can be trained to determine, by visual proximity, whether a given longitudinal image belongs to the first longitudinal zone Z1, the second longitudinal zone Z2, or an uncertainty zone Z? located between zones Z1 and Z2. The goal is to determine the boundaries of the smallest possible zone Z? and to reconstruct a cross-sectional image 58a. Training can be performed on longitudinal images.

[0062] The uncertainty zone Z? is the area where the interface between the first longitudinal zone Z1 and the second longitudinal zone Z2 is located, namely, in this case, the dermo-epidermal junction. Therefore, it is useful to explicitly locate it beforehand to guide the segmentation process. The actual segmentation can then be performed in step 60, based on image 58a and the located zones.

[0063] In step 60, an unsupervised classifier takes as input image 58a or any information representative of this image, for example, indications of the position of the different zones, and determines the position of the dermo-epidermal junction. The classifier used can be of the "zero-cost transfer method" type, meaning that it can adapt to new images and new tasks without prior knowledge. An example of such a classifier is the SAM model (Kirillov, A. et al: Segment anything. ArXiv 2304.02643. 2023) or one of its derivatives (FastSAM, SAM2, etc.). The detection of the second longitudinal zone Z2 allows a rectangle to be automatically positioned in said zone Z2 (i.e., here in the dermis), and the The SAM model extends this rectangle and modifies its shape up to the limit of the area to be detected, namely here the dermo-epidermal junction.

[0064] If needed, during a step 62, the image can be enhanced by image processing, for example based on mathematical morphology operators, typically non-linear operators such as erosion, dilation, opening or closing.

[0065] If necessary, the previously removed layers can be reintegrated in step 64 to obtain, in step 66, a complete and segmented image 66a, which shows the outlines of the different layers being sought. As mentioned previously, the living epidermis is deduced to be the layer located between the dermo-epidermal junction and the stratum corneum.

[0066] Instead of producing an image 66a with highlighted layer outlines as shown in Figure 2, the segmentation process could produce any other information representative of the location and boundaries of the layers, for example, a table of values. The segmented image 66a, or any equivalent representation, can then be used to numerically determine various metrics such as the thickness of the segmented layers, the degree of undulation of an interface between two layers—typically the dermo-epidermal junction—or other measurements useful for dermatological analysis to characterize the condition of the skin or the effectiveness of a treatment applied to the skin.

[0067] The inventors compared the proposed segmentation method to other methods based on supervised training or visual analysis, using skin imaging data from three hundred women of different age and ethnic groups, across several areas of the human face (temple, cheekbone, jawline), and multiple images per area, totaling approximately 2,000 3D images. They determined that the results of the segmentation method were very close to those of previously used methods, while offering considerable time savings, eliminating annotation bias, and increasing generalizability. The proposed segmentation method is therefore reliable, particularly because the unsupervised segmentation is guided by prior processing based on the known structure of the skin layers.

[0068] Figure 3 schematically illustrates a skin cross-section image segmentation system configured to obtain a skin cross-section image; process the skin cross-section image, said processing including edge detection on the image and / or image contrast enhancement; and segment the processed image by an unsupervised classifier to identify one or more skin layers.

[0069] The segmentation system 70 here has the hardware architecture of a computer. It includes, in particular, a processor 72, read-only memory 73, random-access memory 74, non-volatile memory 75, and communication means 76 with an acquisition device 71 (or, alternatively, a database) enabling the segmentation system 70 to obtain a cross-sectional skin image. The segmentation system 70 and the acquisition device 71 are, for example, connected by a digital data bus or a serial interface (e.g., USB (Universal Serial Bus) interface) or a known wireless interface.

[0070] The read-only memory 73 of the segmentation system 70 constitutes a recording medium in accordance with this exposition, readable by the processor 72 and on which is recorded a computer program in accordance with this exposition, comprising instructions for the execution of the steps of a segmentation process described previously with reference to Figures 1 and 2.

[0071] This computer program can equivalently define functional modules of the segmentation system 70 capable of implementing the steps of the segmentation process. Specifically, this computer program can define a module 70A for obtaining a transverse skin image, a processing module 70B, and a segmentation module 70C.

[0072] Although this description refers to specific embodiments, modifications may be made to these examples without departing from the general scope of the invention. Furthermore, individual features of the various embodiments illustrated or mentioned may be combined in additional embodiments. Therefore, the description and drawings should be considered illustrative rather than restrictive.

Claims

DEMANDS

1. A computer-implemented method for segmenting a cross-sectional skin image, comprising: - obtaining (12, 50) the transverse skin image; - the processing (20, 54) of the transverse skin image, said processing including the detection of contours on the image and / or the increase of the contrast of the image; - the segmentation (22, 60) of the image processed by an unsupervised classifier to identify one or more skin layers, in which one or more segmented skin layers are removed (56) from the image and the processing and segmentation steps are repeated.

2. Segmentation method according to claim 1, further comprising a preprocessing (14, 16, 18, 52) of the image comprising at least one of normalization, denoising and averaging of several transverse skin images.

3. A segmentation method according to claim 1 or 2, wherein one or more skin layers are segmented from the more superficial layers of the skin to the deeper layers of the skin.

4. A segmentation method according to any one of claims 1 to 3, wherein the segmentation comprises detecting the location of an interface between two layers by classifying longitudinal skin zones (Z?) comprised between one or more first longitudinal zones (Z1) associated with a first of the two layers and one or more second longitudinal zones (Z2) associated with a second of the two layers, and segmenting the interface on the basis of the detected location.

5. A segmentation method according to any one of claims 1 to 4, wherein the processing includes the removal of image areas showing hairs (P).

6. Segmentation method according to claim 5, wherein said hair-bearing areas (P) are detected on one or more longitudinal skin images (30a).

7. Segmentation method according to claim 5 or 6, wherein said hair-bearing areas (P) are detected by a classifier and / or a thresholding mechanism.

8. A segmentation method according to any one of claims 1 to 7, wherein the cross-sectional skin image is obtained from three-dimensional skin imaging data (12a).

9. A segmentation method according to any one of claims 1 to 8, wherein one or more segmented layers comprise the stratum corneum (SC) and optionally the living epidermis located between the stratum corneum and the dermo-epidermal junction.

10. Segmentation system (70) of a cross-sectional skin image, configured to: - obtain a cross-sectional image of skin; - process the cross-sectional skin image, said processing including edge detection on the image and / or increasing image contrast; - segment the processed image by an unsupervised classifier to identify one or more skin layers, the segmentation system (70) being configured to remove one or more segmented skin layers from the image and repeat the processing and segmentation steps.

11. A program set comprising instructions for carrying out the steps of the processing method according to any one of claims 1 to 9 when said program set is executed by at least one computer or microprocessor.

12. Computer-readable recording medium on which is recorded at least one computer program comprising instructions for carrying out the steps of the processing method according to any one of claims 1 to 9.

Images