|Year : 2019 | Volume
| Issue : 2 | Page : 88-99
Objective assessment of skin repigmentation using a multilayer perceptron
Juan Fernando Chica1, Sayonara Zaputt2, Javier Encalada2, Christian Salamea1, Melissa Montalvo1
1 Interaction, Robotics and Automation Research Group (GIIRA), Universidad Politécnica Salesiana, Cuenca, Ecuador
2 Departament of Dermatology, José Carrasco Arteaga Hospital, Cuenca, Ecuador
|Date of Submission||29-Oct-2018|
|Date of Acceptance||27-Dec-2019|
|Date of Web Publication||24-Jun-2019|
Mr. Juan Fernando Chica
Universidad Politécnica Salesiana, Cuenca
Source of Support: None, Conflict of Interest: None
Background: Vitiligo is a pathology that causes the appearance of achromic macules on the skin that can spread on to other areas of the body. It is estimated that it affects 1.2% of the world population and can disrupt the mental state of people in whom this disease has developed, generating negative feelings that can become suicidal in the worst of cases. The present work focuses on the development of a support tool that allows to objectively quantifying the repigmentation of the skin. Methods: We propose a novel method based on artificial neural networks that use characteristics of the interaction of light with the skin to determine areas of healthy skin and skin with vitiligo. We used photographs of specific areas of skin containing vitiligo. We select as independent variables: the type of skin, the amount of skin with vitiligo and the amount of repigmented skin. Considering these variables, the experiments were organized in an orthogonal table. We analyzed the result of the method based on three parameters (sensitivity, specificity, and F1-Score) and finally, its results were compared with other methods proposed in similar research. Results: The proposed method demonstrated the best performance of the three methods, and it also showed its capability to detect healthy skin and skin with vitiligo in areas up to 1 × 1 pixels. Conclusion: The results show that the proposed method has the potential to be used in clinical applications. It should be noted that the performance could be significantly improved by increasing the training patterns.
Keywords: Artificial neural network, multilayer perceptron, objective assessment, skin repigmentation, vitiligo
|How to cite this article:|
Chica JF, Zaputt S, Encalada J, Salamea C, Montalvo M. Objective assessment of skin repigmentation using a multilayer perceptron. J Med Signals Sens 2019;9:88-99
|How to cite this URL:|
Chica JF, Zaputt S, Encalada J, Salamea C, Montalvo M. Objective assessment of skin repigmentation using a multilayer perceptron. J Med Signals Sens [serial online] 2019 [cited 2020 Sep 30];9:88-99. Available from: http://www.jmssjournal.net/text.asp?2019/9/2/88/259724
| Introduction|| |
The skin is the largest organ of the human body. Its main function is to protect the internal organs from agents that are in the environment. However, the cells that compose it can present disorders in their functions, producing multiple pathologies such as cancer, psoriasis, and vitiligo. Among cutaneous diseases, one of the most frequent pigment pathologies worldwide is vitiligo, which affects between 0.5% and 1.2% of the world population. Although it affects without distinction of sex, race or age, and even some studies have reported that there is a higher incidence of vitiligo in India, Mexico, and Japan.,, Vitiligo is characterized by causing the selective destruction of melanocytes (cells responsible for producing a skin protection pigment called melanin), resulting in the appearance of achromic macules that can change shape, increase in size or be distributed in other areas of the body.,
Vitiligo can be psychologically devastating affecting the mental state of people generating negative feelings such as anxiety and depression,,, fear, and even suicidal thoughts. Although its etiopathogeny is multifactorial, it has been identified that biochemical and cytotoxic changes generated when a person is being subjected to high levels of stress, which becomes the main trigger and exacerbate of this pathology.
The therapeutic treatment for this pathology is carried out from two approaches: the medical and the surgical, including oral treatments, topical, phototherapy, and alternative methods such as relaxation therapies.,, The fundamental objective of these, is, first to stop the progress of vitiligo and then to recover affected skin area by repigment it. The effectiveness of the treatment is evaluated through the analysis of the skin, identifying factors such as color changes (peripheral o perifollicular) and size of the affected area. Nowadays, subjective, semi-objective, and objective methods have been proposed to evaluate the skin affected by vitiligo. Based on the objective methods, high-precision noninvasive support methods using image processing have been developed by taking advantage of the high sensitivity of image sensors installed in conventional digital photographic cameras. Fadzil et al. using an image obtained with a conventional digital camera, they proposed a method to measure the repigmentation of the skin. Using principal component analysis and the technique based on the independent component analysis (ICA) known as FastICA, two images that are based on the melanin and hemoglobin of the skin were obtained. Finally, using image based on melanin, the amount of healthy skin and skin with vitiligo in the photograph was determined. Nurhudatiana, proposed the use of only the blue channel of a red-green-blue (RGB) photograph and with a clustering algorithm called Fuzzy C-Means (FCM) to group the pixels that correspond to healthy skin and skin with vitiligo in two different groups. Once grouped, the group with the lowest centroid value defines the pigmented skin, while the group with the highest centroid value defines the skin with vitiligo. It should be noted that the experiments were carried out using images taken from the internet. In addition to techniques that use RGB photographs, other methods have been developed with the use of spectra diffuse reflection of the skin, models of the skin or other computational algorithms different from FastICA and FCM.,,
The present work focuses on the development of a support tool for the diagnosis and evaluation of the treatment that allows to objectively quantifying the repigmentation of the skin. To this end, we propose a novel method that uses the characteristics from the interaction of light with healthy skin and skin with vitiligo, which then can be captured with a conventional digital photographic camera. The characteristics extracted from the photograph are applied to an artificial neural network (ANN) of multilayer perceptron type (MLP) of supervised learning, which is responsible for determining pixel to pixel the areas affected by the pathology, allowing a quantitative measure of them. In related papers, have been shown that in an RGB image of the skin, the pixels contain information related to the melanin (mostly responsible for the color skin) and hemoglobin but these techniques used the entire photograph. Based on this, we proposed a method that analyzes a photograph pixel to pixel, and to avoid the problems with the generality of the model and the different skin types, we use statistical metrics that give to the MLP information about of the distribution (central tendency and scattering) of the values of the intensity levels of the pixels in all the image. Thereby, the proposed method uses specific information of a single pixel; in addition to the information provided by the entire photograph.
The use of ANNs due to the good results obtained when analyzing medical images,,, have been opted. Among the various types of ANNs, an MLP type because once trained have been chosen, its implementation does not generate excessive computational cost; so, it could be implemented in mobile applications. Besides, we also aimed to evaluate the utility of its nonlinearity characteristics to recognize patterns pixel to pixel in a medical image. The method proposed in this work has been compared with two other methods that exist in the scientific literature to evaluate its performance.,
Interaction of light with skin
The skin is composed of several layers of tissues, in other terms; the stratum corneum (outer layer of the skin), the epidermis, the dermis, and the subcutaneous tissue or hypodermis. When a beam of light hits the skin, only a small fraction is reflected directly (4%–7%), the rest penetrates and follows a path through the first layers of the skin, until it comes back out, or dims by the chromophores (molecule that absorbs electromagnetic energy) of the skin. The predominant chromophores in the dermis and epidermis are oxyhemoglobin (OHb), deoxyhemoglobin (DHb), and melanin.,
The efficiency in the absorption of the electromagnetic energy of the chromophores of the skin is a function of the wavelength of the electromagnetic energy, the OHb has a higher absorption in the bands of 412 nm, 542 nm, and 577 nm, whereas the greater absorption of DHb occurs in the 430 nm and 555 nm bands. Melanin, on the other hand, does not show maximum absorption in the visible light region, nevertheless, increases progressively when approaching the short-wave electromagnetic spectrum. [Figure 1]a shows that in the red spectrum region (600 nm), the absorption of melanin is greater compared to OHb and DHb [Figure 1]b.
|Figure 1: Absorption spectrum of the skinīs predominant chromophores: (a) response profile of Eumelanin, it can be noted that although there is no maximum absorption in the region of visible light, it increases progressively when approaching the electromagnetic spectrum shortwave; Image based on. Absorption spectrum of the skinīs predominant chromophores (b) response profile of oxy-hemoglobin (solid line) and deoxyhemoglobin (dashed line). Image based on|
Click here to view
Thus, when a beam of light interacts with the skin, the information of melanin and hemoglobin (OHb and DHb) is found in the total reflection (resulting light), information that can be recorded in a digital image.
Artificial neural networks
The nervous system of the human body is made up of millions of interconnected nerve cells called neurons. The neuron could be defined as an information processor that is composed of dendrites (input channel), the soma (where the information is processed), and an axon (output channel). On the other hand, the neuron has an excitation threshold (action potential), which is a minimum level that should have a stimulus to activate the neuron and transmits that stimulus to another neuron, in case the stimulus does not beyond the threshold, it will not be transmitted.
If a stimulus is present, the neuron communicates with other neurons (in a process called synapse) and forms connections that allow an adequate response to that stimulus. If the same stimulus occurs again and again, the neurons can improve or inhibit their response in such a way that the connections between them are the most optimal to provide the best possible response to that stimulus; the process of presenting the stimulus over and over again is called training.
Artificial neurons, on the other hand, are formed from a mathematical model that simulates the behavior of biological neurons, both in their function and their way of processing information. [Figure 2] shows a model of the artificial neuron, relating this model to the biological neuron, the set of inputs Xi represents the dendrites, similar to what happens in the synapse in the artificial neuron, it can inhibit or improve its response to mathematically model, by the modification of numerical values called synaptic weights (W). In this way, each of these inputs is multiplied by its respective synaptic weight (Wij) and a weighted sum of these values is performed. In the event that the sum exceeds a certain threshold value (analogous to the excitation threshold), the resulting value is evaluated by an activation function f () (analogous to the action potential) with which the output of the neuron is obtained (Tj). The output of a neuron can be described by the following mathematical function:
The artificial neuron allows to solve simple and linear problems. Thus, when the artificial neurons are interconnected with other neurons, structures called ANNs can be formed to solve complex problems. There are different types of ANN topology, being that the MLP is the most widely used.,, Although the MLP architecture is diverse, this topology consists of three layers of neurons: input (receives data), hidden (processes information), and output layer (provides the output of the network). [Figure 3] shows the architecture of an MLP with a hidden layer where each neuron (represented by a circle) is linked to another through a synaptic weight (W).
There are several types of activation functions within the ANNs. In the case of MLP, the most commonly used are the sigmoidal function (restricts its output to values of 0 and 1) and the hyperbolic tangent functions 1 (restricts its output to values of −1 and 1).
The MLP has the ability to learn by adjusting the value of their synaptic weights through a training phase. In the training phase, input patterns are presented, and the desired output is indicated for each pattern to the network, this process is repeated in a cyclical way. Thereby, when all the training patterns have been presented to the network, a process called epoch is completed. The number of times required for in the training phase depends on the type of problem being treated, and it could range from 1 time to 10,000 or more; depending on whether the network has reached the value of the desired error. The purpose of training is for the network to generalize the information in such a way that it knows how to respond correctly when data are presented for which it was never trained, making it necessary to choose correctly the parameters of the network (number of neurons, activation function, and training data). In this way, problems of overlearning can be prevented (the neural network gives a correct answer only for the training data but not so for new data) or prevent the network from not being able to generalize because there is not enough training patterns.
In case that images are MLP inputs, the training process consists on using input patterns (groups of pixels) as well as inputs to the ANN, and in turn knowing the desired outputs for each particular pattern. After all the input patterns have been presented to the network, what has been defined as an epoch is obtained. This process will be repeated iteratively until the MLP has adjusted its Wij synaptic weights in such a way that the difference (error) between the desired output and the current output is minimal.
The conjugate gradient methods (CGMs) are currently the most popular iterative algorithms when is necessary to solve linear systems “defined-positive” and symmetric equations. Like the standard backpropagation method, the CGM uses the second derivative of the objective function to approach the minimum value in an iterative process, guiding itself in a conjugated direction with respect to the previous steps. In the case of ANNs, the conjugate gradient can be used in updating weights, to produce a response that generates the least possible error in the network. Being W the matrix that represents the synaptic weights, a(k) the matrix of second derivatives (Hessian matrix) andp(k) a function of a(k) that contains the error function that we want to minimize. In this way, the update of synaptic weights would be given by:
To optimize the iterative process, Moller proposed the scaled conjugate gradient (SCG) method, which differs from conventional CGM in two aspects: the first, for the calculation of the error function (key term in the calculation of a[k]), a long search process is not used, but a simple approximation is made. The second, to ensure good performance is used to (k) as a scalar, which implies a regulation of the Hessian matrix (the Hessian matrix may not always be positive definite). The SCG has been shown to be faster than the standard retro-propagation and other CGMs.
| Materials And Methods|| |
The study protocol was approved by the Research Coordination Center of a Hospital in Cuenca Ecuador. Twenty patients from this institution were selected to obtain the photographs used in the training and validation of the proposed method.
Inclusion criteria: Patients who participated in the study signed an informed letter of consent, they have Type II or type III skin according to the Fitzpatrick scale, their age varies between 18 and 55 years, and they did not have any type of burn (sunburn and phototherapy) near the photographed area.
Reflection of light
The areas affected by vitiligo, typically do not contain melanin or are unable to produce it. When a light strikes the affected skin, as there is no melanin and hemoglobin that absorbs part of the electromagnetic energy of light; therefore, the resulting light is attenuated. Nevertheless, in the case of skin with vitiligo, the only absorption that will exist will be that of hemoglobin, causing the intensity of the resulting light not to be attenuated and thus greater than that of healthy skin.
[Figure 4]a and [Figure 5]a show two different types of skin: skin Type II and skin Type III, respectively. In these images, the absence of melanin generates achromic macules due to how the light reflects on the skin. To observe the differences of absorption and reflection of light in the skin in each of the RGB bands, a row of each of the images has been extracted (indicated as a black line on the image) and they have been plotted in [Figure 4]b and [Figure 5]b, respectively.
|Figure 4: Level of reflection of the skin in red-green-blue images: (a) Original skin image. Level of reflection of the skin in red-green-blue images: (b) Information of the intensities of the red-green-blue image obtained from row 126 represented by a horizontal black line in the original image|
Click here to view
|Figure 5: Level of reflection of the skin in red-green-blue images: (a) Original skin image. Level of reflection of the skin in red-green-blue images: (b) Information of the intensities of red-green-blue image obtained from row 202 represented by a horizontal black line in the original image|
Click here to view
In [Figure 4]b, it can be seen that in the area affected by vitiligo there is no light attenuation due to the absence of melanin, so the value of the intensity of the pixels in the RGB bands is higher in this area [approximately from pixel 110–140 in [Figure 4]b. Similarly, in [Figure 5]b, it can be seen that the intensity of the pixels in each of the bands is greater in the area affected by vitiligo.
Although there is a similar pattern of light reflection in both cases, the range of intensity values in clear skin is different from darker skin. However, although the magnitudes are different, differences in the reflection of light can be used by an algorithm to differentiate healthy skin from the skin with vitiligo.
We propose the use of an ANN with an MLP topology to analyze each of the pixels of an image and classify them, either as healthy skin or skin with vitiligo; based on the light reflection pattern. Therefore, at the end of the process, the MLP generates a binary image from the original image, where the pixels with the value of 0 correspond to healthy skin, and the pixels with the value of 1 correspond to skin affected by vitiligo.
The proposed MLP consists of three layers; input layer, hidden layer, and output layer. In the hidden layer a sigmoidal logarithmic activation function is used because its nonlinear response makes it ideal for the classification of this type of patterns. The RGB components of each pixel of the image are used as input characteristic vectors. Due to the difference in the intensity level of light reflected by healthy skin and the skin affected by vitiligo varies depending on the skin tone, two vectors of additional characteristics that remain constant during the analysis of the image has been included to improve the performance of the method; on the one hand, the CV or coefficient of variation,, that indicates how scattered are the magnitudes of the data in the image and the average of the minimums of each row of RGB image that gives us information about the type of skin. The feature extractions are applied to every image, and the method is shown in [Figure 6], on the other hand, [Figure 7] shows the general architecture of the MLP network.
|Figure 6: Featured extraction of images, (a) Input of multilayer perceptron. Featured extraction of images, (b) Desired Output|
Click here to view
|Figure 7: Proposed architecture of multilayer perceptron. It should be noticed that the proposed method analyze the image pixel to pixel|
Click here to view
Training of multilayer perceptron
From the 20 photographs we obtained a total of 2,182,206 (number of pixels of the total images) input patterns, and 2, 182, 206 desired outputs. Whereas, each input patternPis formed by 5 features, the 3 RGB components of each pixel of the original image; plus, the mean of the minimums and the variation coefficient of the photograph. To obtain the desired output T, each photograph has been binarized manually with the help of dermatologists, thus the value of 0 has been assigned to the pixels corresponding healthy skin; and the value of 1 has been assigned to the pixels corresponding skin with vitiligo [Figure 6]b. In the end, each T pattern has two states only, 1 or 0; resulting the input pattern to be P5 × 2 182206 and the desired output is T1 × 2 182206.
Of all the obtained data, 70% was used for training, 15% for cross-validation and the remaining 15% for the evaluation of the generalization and learning of the MLP. Because there is no analytical method to determine the number of neurons in the hidden layer, several tests were performed modifying the number of neurons in the hidden layer. During the training test, 10,000 epochs were used. The smallest error was reached in the 9941 period with a mean square error of 0.1468, both for validation and for testing. The best result was obtained by using 12 neurons in the hidden layer.
To evaluate the performance of the proposed method (MLP) to detect areas of repigmentation in the skin, pattern images were constructed. To do this, samples of the skin color (healthy and with vitiligo) of different patients were selected. It should be noted that the images used in the experimentation are not the same as those used during the training of the MLP network.
The characteristics of the pattern images were selected using an orthogonality based experiment (experiments are organized in an orthogonal table optimally and statistics are applied to analyze their results, the design consists of: Determine the objective of the experiment, determine the experimental factors and define their level, choose the exact form of the orthogonal table and finally place all the previous parameters within the table). Based on these, two experiments were developed. In the first experiment, eight pattern images using modeled skin (Gaussian distribution was used to model healthy skin and skin with vitiligo) were constructed [Figure 8] where: Skin type (based on the Fitzpatrick scale), the amount of skin with vitiligo, and the amount of repigmented skin that may exist in an image were selected as independent variables. On the other hand, in the second experiment, 12 pattern images were constructed [Figure 9], but using real skin where: six samples of six different patients were selected, two patients with skin Type II, two patients with skin Type III, one patient with skin Type II, and one patient with skin Type III, but additionally, there is body hair in both, the healthy skin and the skin with vitiligo, similarity to the first experiment skin type, the amount of skin with vitiligo, and the amount of repigmented skin that may exist in an image were selected as independent variables. The amount of affected skin and the type of skin color modeling was based on a similar study developed by Fadzil et al. The scheme of the first and the second experiment are shown in [Table 1] and [Table 2], respectively.
|Figure 8: Pattern images constructed based on Table 1: (a-h) Test 1–7. It can be noticed that is a modeled skin|
Click here to view
|Figure 9: Pattern images constructed based on Table 2: (a) Test 9, (b) test 10, (c) test 11, (d) test 12, (e) test 13, (f) test 14, (g) test 15, (h) test 16, (i) test 17, (j) test 18, (k) test 17, (l) test 17|
Click here to view
The performance of the method is measured based on the detection rate of healthy skin and skin with vitiligo. To do this, sensitivity (Eq. 3), specificity (Eq. 4), and the F1 Score (Eq. 5) are used as metrics (it indicates the accuracy of a method with respect to its sensitivity and precision). In the case of the F1-Score, a value of 1 equals 100% accuracy, this value can be calculated using:
Where TP (true positives) represent the skin pixels with vitiligo correctly detected, FP (false positive) represents the healthy skin pixels detected as skin with vitiligo, and FN (false negative) represents the skin pixels with vitiligo detected as healthy skin. In addition, two similar work methods were used (method based on ICA and method based on FCM) to compare their results with respect to the proposed method.
| Results and Discussion|| |
[Table 3] shows the results of the methods in the first experiment. In the first four tests (Test 1 to Test 4), what is sought is to evaluate the performance of the methods to identify areas of depigmentation that correspond to skin affected with vitiligo.
The results indicate that the proposed method and the ICA method are both capable of correctly detecting the pixels that either correspond to healthy skin or skin with vitiligo. In the first four tests, the three methods obtained a score of 100% with respect to their sensitivity, with the exception of the test 4, where the ICA method achieved 96% of sensitivity. With respect to its specificity, in Test 1 and Test 2, the method of ICA obtained better results than the proposed method, however, said proposed method also obtained a good result as it reached 99.62% for both tests; that is to say with regard to the method of ICA a difference of 0.35% and 0.04%, respectively. On the other hand, in Tests 3 and 4, the proposed method was the better score obtained from the three applied methods; as in being 100% accuracy to detect the pixels of healthy skin and skin with vitiligo.
It should be noted that the low results obtained in the F1-score were due to the fact that there is a single value (1 pixel) corresponding to a true positive so that when said method detects a false positive or false negative, it greatly diminishes the score on this metric and not necessary shows low performance. If the scores obtained by both methods are to be observed, these are very acceptable (around 99% accuracy); however, due to the false positives and false negatives detected their F1-score value is low. The FCM method did not obtain good results in any of the cases, being that in its best test it obtained a result of 51.14% of specificity. Its low performance in these tests was due to the fact that being a clustering algorithm it divides the photo into two groups, in this case, having been only one (1 × 1) or 25 (5 × 5) pixels that contain skin with vitiligo, the algorithm did not consider them to be a reference in any of two groups (healthy skin and skin with vitiligo). The results of the three methods subjected to Test 1 are shown in [Figure 10], from which it can be noticed the low performance of the FCM method in contrast with the results of other methods.
|Figure 10: Responses of the three methods to the test 1: (a) Pattern image, (b) desired response, (c) independent component analysis method, (d) Fuzzy C-Means method, (e) proposed method|
Click here to view
The following tests (Test 5–Test 8) look to evaluate the performance of the methods to detect correctly an area of repigmentation that could exist in an area affected with vitiligo, and simultaneously its accuracy to detect healthy skin and skin with vitiligo as such. For these tests, the three methods obtained quite similar results; furthermore, the FCM method was capable of identifying each type of skin due to the fact that the area with affected skin was greater in the first tests, allowing it to identify the pixels that corresponded to each type of skin.
The F1-Score allows us to have a general vision of the accuracy of a determined method, in these tests the amount of true positives are significant (around 2500) with respect to the FP and FN of the image; and thereby the metric of the F1-score may be valid to give a benchmark of performance of the methods. Based on this parameter, among these tests, the proposed method was the most accurate being that it obtained the highest F1-score; obtaining an average of the 4 tests, an F1-Score of 0.9851 while the ICA method and the FCM method achieved an F1-score of 0.9558 and 0.9612, respectively. In some of the tests, the proposed method did not obtain the best result in regards to its sensitivity or its specificity but obtain the highest score in the F1-Score. This fact must be considered for future references due to the false positives in tests could be confused as skin with vitiligo, and thus show a greater area of skin as being affected than what in reality is. The results of the three methods in Test 8 are shown in [Figure 11], from which it can be noticed that methods detect correctly the skin with vitiligo but also some false positives.
|Figure 11: Responses of the three methods to the test 8: (a) Pattern image, (b) desired response, (c) independent component analysis method, (d) Fuzzy C-Means method, (e) proposed method|
Click here to view
[Table 4] shows the results of the methods in the second experiment, using real skin samples. It should be noted that some areas of the skin of certain individuals, it tends to generate a shine peculiar of each person and skin type being that in these cases the proposed method and the other methods confuse this shine as skin with vitiligo. In the test where only exist one pixel with vitiligo the FCM method still not obtain good results, similar to the results in the first experiment in this kind of pattern images.
It can be noted that the proposed method and the ICA method detect various false positives (could be caused by shine skin) but it also detects correctly the 1 pixel corresponding to skin with vitiligo. Although FP and FN were detected in some tests, the proposed method obtained the best results in contrast to the other tested methods and due to in this tests are using real skin the proposed method shows potential to be used in clinical applications. This fact can be observed in [Figure 12] (the results of the methods in the Test 9 is shown).
|Figure 12: Responses of the three methods to the test 9: (a) Pattern image, (b) desired response, (c) independent component analysis method, (d) Fuzzy C-Means method, (e) proposed method|
Click here to view
The last four tests (Test 17–Test 20), has the particularity that the real skin samples contain healthy skin, skin with vitiligo an also body hair. The FCM method incorrectly detects lots of pixels, that is, the fact of its lower metrics in [Table 4]. However, It is interesting to noted that the presence of body hair not affect significantly the response of the proposed method and the ICA method, though the proposed method obtained better results, except in the test 20; in this test due to the characteristic of the skin some areas has a kind of brightness causing the false positive detection. [Figure 13] shows the results of the methods in the test 20, where can be observed the mentioned above.
|Figure 13: Responses of the three methods to the test 20: (a) pattern image, (b) independent component analysis method, (c) Fuzzy C-Means method, (d) proposed method|
Click here to view
| Conclusion|| |
In this paper, an MLP and the characteristics of the light interaction with the skin were used to determine the pixels corresponding to healthy skin and skin with vitiligo from an RGB image. Its results were contrasted with two methods of similar works. It is noteworthy that the tests tried to simulate different scenarios that could happen in a real case.
We analyze the result of the methods based on three parameters (sensitivity, specificity, and F1-Score) to have a global vision of their performance. The proposed method demonstrated the best performance of the three methods, and it also showed its capability to detect healthy skin and skin with vitiligo in areas up to 1 × 1 pixels. In tests with real skin, the proposed method detected various false positives; however, among all the methods, it obtained the best possible results. In the pattern images that contain body hair, it can be observed that the proposed method has a good response according to the other methods, to different conditions of the skin of the skin, in this kind of images, the body hair did not represent problems, but areas on the skin that reflect too much brightness in the photograph is a factor to consider due to could be a limit in any method that uses images. However, related to clinical applications the proposed method present several advantages based on the obtained results. It should be noted that some photographs could have very thick body hair, in this case, the proposed method could not respond accurately, but this could be due to the thickness of the body hair that could be covered the skin by them resulting in an erroneous response. The relevance of this method lies in the fact that ANNs only require huge computational cost during its training; being that a trained network could be easily used in mobile health, and if it is required, the performance of the proposed method would be significantly improved by increasing the training patterns. The next step in our research, considering the problems due to shine (detecting false positives) and thick body hair, will be to develop a method based on machine learning algorithms that are robust to the brightness that may exist on the skin and it can be consider the use of a preprocessing technique to remove the thick body hair, if its required.
Providing the physicians and patients an objective method to evaluated pigmentary disorders such as vitiligo will help doctors and patients to make smarter and faster decisions during the evaluation process, helping to decreases the stress caused by the lengthy treatment of patients.
The authors would like to thank the patients who participated in this study. Authors greatly acknowledge the Universidad Politecnica Salesiana and the José Carrasco Arteaga Hospital by the openness and support to carry out this research. We would like to acknowledge Flor Urgiles, from the Dermatology Department, for her assistance with the patients.
Financial support and sponsorship
Universidad Politecnica Salesiana.
Conflicts of interest
There are no conflicts of interest.
| Biographies|| |
Juan Fernando Chica Ortiz, received his Electronic Engineer degree from the Universidad Politecnica Salesiana, Cuenca, Ecuador, in 2017. He was a member of the Artificial Intelligence and Assistance Technologies Research Group (GIIATA) of the Universidad Politecnica Salesiana until 2016. From 2018 hes a master student on a artificial intelligent in the Universidad Internacional de la Rioja. Currently, he is a researcher in the Interaction, Robotics, and Automation Research Group (GIIRA) of the Universidad Politecnica Salesiana. His areas of interest are the processing of voice signals and images, and applications of artificial intelligence algorithms.
Sayonara Zaputt, obtained her degree in Doctor of Medicine and Surgery at the State University of Cuenca of Ecuador in 2001, her degree as a Dermatologist is obtained at the University of El Bosque in Bogotá - Colombia in 2010, currently working as a dermatologist in José Carrasco Arteaga Hospital and University Hospital of Rio, she is a professor of the Chair of Dermatology at the University of Cuenca. His areas of interest include advances in prevention and therapeutics in inflammatory skin diseases and pigmentation disorders.
Javier Encalada, obtained his degree of Doctor of Medicine and Surgery from the University of Cuenca of Ecuador in 2000, obtained his degree in Dermatology at the University of Buenos Aires - Argentina in 2006, currently working as a dermatologist and academic coordinator in the department of medicine at the Hospital José Carrasco Arteaga, Cuenca - Ecuador. His areas of interest include advances in prevention and therapeutics in inflammatory skin diseases and pigmentation disorders.
Christian Salamea, obtained his Electronic Engineer degree in 2004, from the Universidad Politecnica Salesiana of Ecuador and his Master's degree at the Universidad Politécnica de Catalunya in 2009. From 2004 to the present he works as a professor at the Universidad Politecnica Salesiana in related areas with Electronic Engineering. In 2018 he obtain a PhD in the Speech Technology Group of the Universidad Politécnica de Madrid. Currently, his work is focused on the recognition of language, particularly in areas related to phonotactic systems and neural networks.
Melissa Montalvo, obtained her degree in Electronic Engineering from the Universidad Politecnica Salesiana of Ecuador in 2017. Since 2016, she is currently a researcher in the Interaction, Robotics, and Automation Research Group (GIIRA) of the Universidad Politecnica Salesiana. Her areas of interest include the processing of biomedical signals, artificial intelligence systems, and deep learning networks.
| References|| |
Yaghoobi R, Omidian M, Bagherani N. Vitiligo: A review of the published work. J Dermatol 2011;38:419-31.
Alghamdi KM, Kumar A, Taïeb A, Ezzedine K. Assessment methods for the evaluation of vitiligo. J Eur Acad Dermatol Venereol 2012;26:1463-71.
Ezzedine K, Eleftheriadou V, Whitton M, van Geel N. Vitiligo. Lancet 2015;386:74-84.
Lakhani DM, Deshpande AS. Various treatments for vitiligo: Problems associated and solutions 2014;4:101-5.
Alikhan A, Felsten LM, Daly M, Petronic-Rosic V. Vitiligo: A comprehensive overview part I. Introduction, epidemiology, quality of life, diagnosis, differential diagnosis, associations, histopathology, etiology, and work-up. J Am Acad Dermatol 2011;65:473-91.
Vrijman C, Hosseinpour D, Bakker JG, Wolkerstorfer A, Bos JD, van der Veen JP, et al.
Provoking factors, including chemicals, in Dutch patients with vitiligo. Br J Dermatol 2013;168:1003-11.
Njoo MD, Westerhof W. Vitiligo. Pathogenesis and treatment. Am J Clin Dermatol 2001;2:167-81.
Parsad D, Dogra S, Kanwar AJ. Quality of life in patients with vitiligo. Health Qual Life Outcomes 2003;1:58.
Schmid-Ott G, Künsebeck HW, Jecht E, Shimshoni R, Lazaroff I, Schallmayer S, et al.
Stigmatization experience, coping and sense of coherence in vitiligo patients. J Eur Acad Dermatol Venereol 2007;21:456-61.
Sampogna F, Raskovic D, Guerra L, Pedicelli C, Tabolli S, Leoni L, et al.
Identification of categories at risk for high quality of life impairment in patients with vitiligo. Br J Dermatol 2008;159:351-9.
Amer AA, Mchepange UO, Gao XH, Hong Y, Qi R, Wu Y, et al.
Hidden victims of childhood vitiligo: Impact on parents' mental health and quality of life. Acta Derm Venereol 2015;95:322-5.
Papadopoulos L, Bor R, Legg C. Coping with the disfiguring effects of vitiligo: A preliminary investigation into the effects of cognitive-behavioural therapy. Br J Med Psychol 1999;72 (Pt 3):385-96.
Falabella R, Barona MI. Update on skin repigmentation therapies in vitiligo. Pigment Cell Melanoma Res 2009;22:42-65.
Felsten LM, Alikhan A, Petronic-Rosic V. Vitiligo: A comprehensive overview part II: Treatment options and approach to treatment. J Am Acad Dermatol 2011;65:493-514.
Fadzil MH, Norashikin S, Suraiya HH, Nugroho H. Independent component analysis for assessing therapeutic response in vitiligo skin disorder. J Med Eng Technol 2009;33:101-9.
Hyvärinen A. Fast and robust fixed-point algorithms for independent component analysis. IEEE Trans Neural Netw 1999;10:626-34.
Nurhudatiana A. A computer-aided diagnosis system for vitiligo assessment: A segmentation algorithm. In: Intan R, Chi CH, Palit HN, Santoso LW, editors. Intell Era Big Data. Berlin, Heidelberg: Springer Berlin Heidelberg; 2015. p. 323-31.
Prince S, Malarvizhi S. Analysis of diffuse reflectance spectra of various skin conditions by principal component method. In: Biomed Pharm Eng. 2009 ICBPE09 International Conference On. IEEE; 2009. p. 1-4.
Hani AF, Nugroho H, Shamsudin N, Baba R. Melanin determination using optimised inverse Monte Carlo for skin – Light interaction. In: 2012 4th
International Conference on Intelligent and Advanced Systems. IEEE. ICIAS; 2012. p. 314-8.
Hassanpour R, Shahbahrami A, Wong S. Adaptive Gaussian mixture model for skin color segmentation. World Acad Sci Eng Technol 2008;41:1-6.
Rani N, Nalam M, Mohan A. Detection of skin cancer using artificial neural network. Int J Innov Adv Comput Sci 2014;2.
Choudhari S, Biday S Artificial neural network for skin cancer detection. International Journal of Emerging Trends & Technology in Computer Science (IJETTCS) 2014;3:147-53.
Wallace VP, Bamber JC, Crawford DC, Ott RJ, Mortimer PS. Classification of reflectance spectra from pigmented skin lesions, a comparison of multivariate discriminant analysis and artificial neural networks. Phys Med Biol 2000;45:2859-71.
Rubegni P, Burroni M, Cevenini G, Perotti R, Dell'Eva G, Barbini P, et al.
Digital dermoscopy analysis and artificial neural network for the differentiation of clinically atypical pigmented skin lesions: A retrospective study. J Invest Dermatol 2002;119:471-4.
Anderson RR, Parrish JA. The optics of human skin. J Invest Dermatol 1981;77:13-9.
Kollias N, Stamatas GN. Optical non-invasive approaches to diagnosis of skin diseases. J Investig Dermatol Symp Proc 2002;7:64-75.
Stamatas GN, Zmudzka BZ, Kollias N, Beer JZ. Non-invasive measurements of skin pigmentation in situ
. Pigment Cell Res 2004;17:618-26.
Durdu A, Cetin H, Komur H. Robot imitation of human arm via artificial neural network. In: Mechatron-Mechatronika ME 2014 16th
International Conference On. IEEE; 2014. p. 370-4.
Nolte J. The Human Brain: An Introduction to its Functional Anatomy. lsevier Canada: Mosby; 2002.
Orhan U, Hekim M, Ozer M. EEG signals classification using the K-means clustering and a multilayer perceptron neural network model. Expert Syst Appl 2011;38:13475-81.
Gardner MW, Dorling SR. Artificial neural networks (the multilayer perceptron) – A review of applications in the atmospheric sciences. Atmos Environ 1998;32:2627-36.
Ceballos-Magaña SG, de Pablos F, Jurado JM, Martín MJ, Alcázar Á, Muñiz-Valencia R, et al.
Characterisation of tequila according to their major volatile composition using multilayer perceptron neural networks. Food Chem 2013;136:1309-15.
Shewchuk JR. An Introduction to the Conjugate Gradient Method without the Agonizing Pain. Carnegie-Mellon University. Department of Computer Science; 1994.
Møller MF. A scaled conjugate gradient algorithm for fast supervised learning. Neural Netw 1993;6:525-33.
Fitzpatrick TB. The validity and practicality of sun-reactive skin types I through VI. Arch Dermatol 1988;124:869-71.
Gravetter FJ, Wallnau LB. Statistics for the Behavioral Sciences. Canada: Cengage Learning; 2016.
Abdi H. Coefficient of variation. Encyclopedia of Research Design. The University of Texas at Dallas, Thousand Oaks; 2010. p. 169-71.
Everitt B, Howell DC, editors. Encyclopedia of Statistics in Behavioral Science. Hoboken, N.J: John Wiley & Sons; 2005.
Pei Z, Zhang Y, Lin Z, Zhou H, Wang H. A Method of Image Processing Algorithm Evaluation Based on Orthogonal Experimental Design. IEEE; 2009. p. 629-33.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6], [Figure 7], [Figure 8], [Figure 9], [Figure 10], [Figure 11], [Figure 12], [Figure 13]
[Table 1], [Table 2], [Table 3], [Table 4]
|This article has been cited by|
||Development of an algorithm to train artificial neural networks for intelligent decision support systems
| ||Oleg Sova,Oleksandr Turinskyi,Andrii Shyshatskyi,Volodymyr Dudnyk,Ruslan Zhyvotovskyi,Yevgen Prokopenko,Taras Hurskyi,Valerii Hordiichuk,Anton Nikitenko,Artem Remez |
| ||Eastern-European Journal of Enterprise Technologies. 2020; 1(9 (103)): 46 |
|[Pubmed] | [DOI]|
||Using multi-layer perceptron with Laplacian edge detector for bladder cancer diagnosis
| ||Ivan Lorencin,Nikola Andelic,Josip Španjol,Zlatan Car |
| ||Artificial Intelligence in Medicine. 2019; : 101746 |
|[Pubmed] | [DOI]|