=Paper=
{{Paper
|id=Vol-3338/ICCS_CVMLH_01
|storemode=property
|title=Human Skin Disease Detection using MLXG Model
|pdfUrl=https://ceur-ws.org/Vol-3338/ICCS_CVMLH_01.pdf
|volume=Vol-3338
|authors=Dhruv Dodia,Hemangi Jakharia,Ritik Soni,Shwetambari Borade,Nilakshi Jain
}}
==Human Skin Disease Detection using MLXG Model==
https://ceur-ws.org/Vol-3338/ICCS_CVMLH_01.pdf
Human Skin Disease Detection using MLXG model
Dhruv Dodia a, Hemangi Jakharia a, Ritik Soni a, Shwetambari Borade a, Nilakshi Jain a
a
Shah & Anchor Kutchhi Engineering College, Chembur, Mumbai, Maharashtra, India
Abstract
Human skin disease detection deals with the recognition of skin type and the skin disease in a
given image. Edges, textures, and colour pixels are often used to detect skin diseases. These
features are invariant and are fast to process. A variant Human Skin Disease Detection model
(MLXG) is put forward in this paper. The three main parameters which help to detect skin
diseases are the edges, the RGB (Red, Green, Blue) colour model of the image, and the
texture of the skin. The goal of the MLXG model is to improve the current human skin
disease detection models in terms of accuracy and to work on different skin diseases. The
model not only considers the above parameters but also uses Machine Learning to improve
the accuracy.
Keywords 1
Skin Disease, Machine Learning, edges, image processing, RGB, dry, acne scars, vitiligo,
warts, alopecia areata, oily, acne, MLXG.
1. Introduction
According to analysis human skin diseases are ranked fourth common disease affecting the world's
⅓ rd population [18]. In a study, a total count of 2701 population was engaged amongst which 1662
entrants i.e. 64.5% had at least one skin abnormality [19]. Human Skin is an important part of the
human body which consists of three layers. Human Skin Disease accrued with ages. Frequency of
increase of human skin disease in males (72.3%) than in females (58.0%)[19]. Skin can also be
affected due to its skin type whether it is oily or dry skin. The diagnosis of skin disease is important
and should not be ignored.
Clinical examination showed that just about a simple fraction of the affected entrants were
unfamiliar with their aberrant skins. Human Skin diseases come up in a variety of forms, lack and
inconsistent availability of dermatologists, need for timely diagnosis to prevent long term effects to
the skin, calls for computer-aided diagnosis. Human skin disease detection is the process of finding
the skin type and the infected part of the skin in a given image provided by the user. Human skin
disease detection system uses a range of image processing, convolution neural network, and machine
learning algorithms. The primary feature for skin disease recognition is the edges, but edges cannot be
the only factor for the same, as there are variations in the skin types according to different conditions
such as geographical conditions, climatic conditions, etc.
The following factors should be considered for determining skin diseases:
1. Varying skin tone with respect to different races will affect the RGB pixel values of the
image.
2. Individual characteristics such as age and gender also affect the contours of the image.
CVMLH-2022: Workshop on Computer Vision and Machine Learning for Healthcare, April 22 – 24, 2022, Chennai, India.
EMAIL: shwetambari.borade@sakec.ac.in (Shwetambari Borade)
ORCID: 0000-0001-7547-6351 (Shwetambari Borade)
©️ 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
1
3. Other factors such as watermark, shadow, blurred image, background colour affect the
contours of the image.
2. Literature Survey
Many Authors and Researchers have imposed Image Processing Techniques and Different
Machine learning algorithms to detect the type of skin and relate to that type of skin disease. We have
reported some techniques which are as follows.
In [1], a system is proposed, image processing technique 2-D Wavelet Transform is used as a
feature extractor from the images, and the detection of disease of the processed images has been done
using Convolutional Neural Network (CNN). They have tested three diseases and have good
accuracy.
The author of [2] applied the method Edge Detection using Gaussian Filter Adaptive Thresholding,
K-Means clustering. Segmentation algorithms are used. Chickenpox, eczema, psoriasis, ringworm are
the diseases detected.
The work [3] of Acne, Amyloidosis, Cherry Angioma, Eczema Lids, Halo-Nevus are the diseases
detected in this technique. KNN is the algorithm used here. The image segmentation is done by the
active delineation method. The feature extraction is done using a mathematical equation, then the
classification of diseases takes place. The image will be grayscale converted, cropped and noise
filtered. The accuracy of this method is 81.92%.
Similarly, [4] the input image is pre-processed by resizing it to the size fixed for the dataset.
Feature extraction was proposed from a pre-trained convolutional neural network. A support vector
machine does the classification. This model detected 3 types of skin with a 100% accuracy rate.
Dataset has 80 images.
Therefore, [5] The processing of an image contains image filtering, image rotation & image
segmentation. In feature extraction, 3 types of human skin diseases are detected. They are given as
follows: herpes, dermatitis, and psoriasis. The average accuracy of this model is 87%. 90 images are
used in the dataset.
According to the observation from Table I that shows the comparison of current work for human
skin disease detection, it has been noticed that the MLXG model that uses the XGBoost algorithm as a
classifier is not yet explored. In this paper, we have illustrated a machine learning system for human
skin disease detection on images using different techniques such as machine learning and deep
learning
Table 1: Comparison Table of Various Survey Papers
Sr Year Paper Disease Diseases Accuracy Method Dataset
no. Cited count
1 2020 [3] 5 • Acne 81.92% KNN ------
• Amyloidosis
• Cherry-Angioma
• Eczema
• Lids
• Halo-Nevus
2 2019 [4] 3 • Melanoma 100% AlexNet - DeepCNN, 80
SVM images
2
• Eczema
• Psoriasis
3 2019 [6] 8 • Acne 88% VGG16 – DeepCNN 6000
images
• Actinic
• keratosis
• Angioedema
• Blepharitis,
• Granuloma facial
• Pityriasis Alba
• Rosacea,
• Vitiligo
4 2019 [2] 4 • Eczema, ----- Kmeans -----
• psoriasis
• chicken pox,
ringworm
5 2018 [5] 3 • Herpes, 85% GLCM - Features 90
Dermatitis, SVM - Classification images
Psoriasis
6 2020 [1] 3 • Psoriasis, ----- 2D Wavelet -----
Transform
• Lichen
algorithm: feature
• Planus, extraction
• Pityriasis convolutional neural
network (CNN):
• Rosea Classification
7 2019 [7] 2 • malignant, 98% GLCM, SVM ------
• melanoma or
seborrheic
• keratosis.
8 2019 [8] 2 • Acne 96% Hessian Matrix 400
• Boils
9 2019 [9] 2 • malignant 85% KNN, SVM -----
• melanoma or
seborrheic
• keratosis.
10 2020 [10] 5 • Nummular 73% CNN 3000
• eczema
• lichen simplex
• stasis
3
• dermatitis
• ulcers
11 2018 [11] 4 • Acne, 86% KNN, MSIN -----
• Psoriasis
• Melanoma
• Heat Rash
Figure 1: Disease selected for Human Skin Disease Detection System
3. Methodology
3.1. Dataset
The dataset itself means collection of data. The diseases which we have selected for our model are
acne, acne scars, vitiligo, alopecia areata, warts shown in fig 1. The dataset has been generated
manually using google images, we have also used images from Dermnet. The images used to build the
dataset are free from watermark, the pixels of the images are in the form of RGB, the background and
blurred part has been cropped. The count of images present in the dataset is shown in Table 2.
The MLXG model, images are divided in the ratio of 80:20 where 80% is used to train and the rest
20% is used to test.
Table 2: Image count in dataset
Classification Task Number of images
Acne 58
Acne Scars 48
Vitiligo 51
Alopecia Areata 55
Warts 45
4
Figure 2: Proposed MLXG Methodology for classifying human skin disease
3.2. Approach
The proposed Machine Learning with XGBoost on microscopic examination of human skin
diseases detection in brief. The MLXG model is classified into two prime parts, the first one is the
feature extraction and the second one is the classifier. The initial step taken in the architecture is pre-
processing the images for uniformity in the images. The pre-processing of images also includes
resizing and resampling. The above-mentioned procedure i.e. data pre-processing is taken to maintain
the uniformity and the quality of the images before passing them further into the Feature extraction or
a machine learning model. The key features from the image are extracted by the network named
VGG16. VGG16 executes an important role in pointing out the features from the images passed after
pre-processing and also explores all the pixels of an image in depth. The final step of the MLXG
model trains all the extracted features from the images using the XGBoost classification algorithm.
The XGBoost algorithm is scalable and incisive to train the MLXG model to distinguish different
human skin diseases. Fig 2 demonstrates the workflow of the proposed Machine Learning with
XGBoost (MLXG) for the Human skin diseases detection expert system.
3.3. Pre-processing and data augmentation
Image processing plays an important role in the images before passing the images in the model.
The randomly downloaded images are collected and further resampled and rearranged. To load the
images, we have used the OpenCV library. Later we have resized the image by 224 * 224 * 3 three
are the channels that say the image is in the RGB form. The image is resized to reduce the calculation
5
of dimensions of each pixel for the feature extractor. The dataset is divided into 80% and 20%,
training and testing respectively.
Data augmentation is a technique in pre-processing for small datasets. A challenge for human skin
disease detection is its limited dataset so we have applied data augmentation to generate a model with
high accuracy and to avoid overfitting. The objective to use data augmentation is not only to increase
the dataset but also to formalize and resolve the disparity. This process will also strengthen the
MLXG model.
3.4. Extraction of features
As extraction of features is an important task over which, the performance of classification of skin
diseases depend. So, we have selected a deep learning perspective for the model to use each pixel of
the image and automatically finds out the features from an image provided. The convolution conducts
signal processing operations that may be calculated in an easy manner as discrete spatial processing
operations, on the basis of principle of Convolution Neural Network (CNN). To implement the model,
we need a Training model and a Classifier Model.
• Training from Scratch
In order for this strategy to be accurate and dependable, it will require a large number of
histological breast pictures to be input. This technique will take more time and effort to achieve since
it necessitates defining and fine-tuning the necessary parameters to reach the optimal results. The
learning rate, number of layers, kind of convolutional layer, and other hyper parameters are all
included in this hyper parameter. Aside from the difficult issue of tuning the parameter, training also
necessitates high-performance GPU processing capability.
• Transfer Learning:
By transferring information learnt from a source domain, such as the dataset from ImageNet with a
huge quantity of data, to the model targeted, we may solve issues like overfitting and create a more
generic deep learning model. In a deep learning model, these models that are pre-trained can give
adequate information and helps in the preparation of the limited histology dataset. This method aims
on using the information gained from a group of data samples and apply it to any future samples that
are not included in the data. The model will be able to fetch all of the learnt features and knowledge to
generate predictions in future for fresh samples in this manner. Learning provides benefits such as
accelerating network convergence, lowering processing power, and improving network performance.
As a result, transfer learning is a more effective technique than training the model from the
beginning for extraction of features from an input by just randomly initialising weights. As a result,
throughout this process, we used pre-trained models. Many deep learning models have enhanced the
CNN model for detecting skin diseases, such as:
1. VGG16 Feature Extraction
For this system, VGG16 (Visual Geometry Group) was selected as the pre-trained feature
extraction model. The VGG16 is improvised because it has less hyperparameters and is a better
version of AlexNet. The VGG16 consists of convolution + ReLU, maximum pooling, and full
connection + ReLu. The input size of the image passed in the first level of convolution is a fixed size
224 * 224 * 3 (RGB channels). The image then goes further through the stack of convolution layers,
with the filter size set to 3 * 3 and the convolution step size set to 1 pixel. Spatial padding for
convolutional layer inputs is such that spatial resolution is maintained after convolution, so 3 * 3
convolutional layer padding is 1 pixel.
6
Figure 3: VGG16 Model [14]
Spatial pooling will continue to be performed using the Max-Pooling layer. Maximum pooling is
performed at 2 * 2 pixels using Stride 2 while keeping the padding the same. Maximum pooling can
be derived using
( N + 2 P - F / S ) + 1.
Figure 4: VGG16 Architecture [14]
Therefore, this is followed by three Fully Connected layers that follow the stack of convolution
layers. The first two fully connected layers contain 4096 channels and the last layer contains 1000
channels. As shown in Figure 4, the VGG16 architecture contains 16 layers. Figure 5 shows the
feature extraction by VGG16.
Figure 5: Training images using VGG16
2. Extreme Gradient Boosting (XGBoost) Classifier
It is a brand-new tree-primarily based totally set of rules that has these days emerge as famous in
facts class and has established to be an enormously powerful facts class method. XGBoost is an
enormously scalable stop-to-stop tree boosting machine utilized in system mastering for class and
regression tasks. XGBoost is extreme, and that means that it is a big machine learning algorithm with
a great number of parts. XGBoost was designed to be used with large, complicated data sets. We are
training data consisting of five different skin diseases. From fig 7 green dots indicate that the
classified skin disease was true and the red dots indicate that the classified skin disease was false. The
first step in adapting XGBoost to your training data is to make an initial prediction. The prediction
can be anything. In our project, the probability of observing skin disease predictions in training data is
0.5 by default, regardless of whether XGBoost is used as a regression or classification.
7
Figure 6: Skin Diseases
From fig 6 we can illustrate the initial prediction by adding a y-axis to our graph to represent the
probability of classification. Drawing a thick line at 0.5 to represent a 50% chance that the
classification is correct. The two green dots we will move up as it represents true classification of
skin disease and the probability is one. Similarly, the two red dots we will leave them at bottom where
the probability is 0 because it represents false classification. The residual, or difference between the
observed and predicted values, shows how good the initial prediction is.
Figure 7: Probability of Classification of Skin Diseases.
The formula to calculate the similarity scores is given in fig 6. Lambda is regularization parameter.
In the next step we tried to fit an XGBoost Tree to the Residuals. Each tree begins off evolved out
as a lone leaf and all the residuals go to the leaf. So now we need to calculate the similarity score for
the leaf. As we do not square the residuals before we add them together, they will cancel each other
out so the similarity score = 0. Next, we need to divide the similar residuals into two groups to
determine if they can be grouped better. So, our next step is to split. As we chose the threshold,
classification < 15 because 15 is the average value between last observations. Thus, the three residuals
with classification < 15 go to the leaf on the left and the one residual with classification > 15 moves to
the leaf on the right side. To calculate the similarity score for the three residuals that ended up in the
leaf on the left. So, we plug in the three residuals into the numerator (-0.5, 0.5, 0.5) and since we are
building the first tree, the Previous Probability refers to the prediction from the initial leaf. So, let's
keep lambda 0.
Thus, the similarity score = 0.33. Similarly, to calculate the similarity score for the leaf on the right is.
Now calculate the Gain.
8
No other threshold gives us a larger Gain value and that means classification < 15 will be the first branch in
our tree. The least number of Residuals in every leaf is determined by calculating cover. Cover is the
denominator of the Similarity Score minus lambda. Gain is associated with the lowest branch and a number we
pick for gamma. The Output value is given by the formula.
Lambda is the regularization parameter and when lambda > 0, it results in greater pruning, by reducing the
similarity scores, and smaller output values for the leaves. Figure gives the idea of tree built and shows the
XGBoost tree.
Figure 8: Tree using XGBoost
Figure 9: Analysis of Average Result of Classification task
9
Table 3: Feature Extraction using VGG16 and Range of the values
Acne Acne Scars Alopecia Areata Vitiligo Warts
Image
of
features
extracte
d using
VGG16
Minimu
-123.68 -123.68 -123.68 -123.68 -80.939
m Value
Maximu
139.061 151.061 133.061 61.06 109.061
m Value
4. Experimental Outcome
In this research, we are focusing to develop a Human Skin Disease Detection using Machine
Learning algorithms. The proposed model is implemented using Python 3.8.5. OpenCV library is used
for the implementation of image processing. The experiment is run on Spyder IDLE. Figure 10
outlines the proposed method.
Figure 10: Overview of the Proposed model
a. Configuration
We used the initial weights trained on the ImageNet dataset to train the VGG16 model on the
proposed system. For fine tuning, image resize and augmentation are implemented to stabilize the
dataset. The image will be converted into BGR. All images used in dataset are resized with the
targeted values 224 * 224 pixels. Further, the image will be converted into an n-dimensional array
using NumPy. The pre-trained model will extract the features from the images. Label Encoder has
been used to encode the labels of the images. As we are working with pre-trained weights so we have
made loaded layers as non – trainable. So, the trainable parameter will be zero once the feature has
10
been extracted by VGG16. The features are extracted from the image and the image is reshaped into
its original form. XGBoost has been used as a classifier algorithm because the optimization of
XGBoost is faster than any other algorithm and it will classify skin diseases.
Figure 11: Confusion Matrix of model MLXG
b. Results
To display the effectiveness of human pores and skin sickness detection, it's far crucial that we
compare our method to test and to recognize the overall performance of the gadget to test if the
underlying troubles are solved or not. Common evaluation for the diagnose are Precision, Accuracy,
F1-measure, and Recall. Training and testing of the final proposed MLXG model was done on around
1000 images of skin disease. Fig 5 illustrates the extraction of features in different layers using the
VGG16 model. The values for each are detected from minimum value till maximum value shown in
Table 3.
Figure 11 shows the confusion matrix for skin disease detection. A confusion matrix is a matrix
commonly used to give an account about the execution and efficiency of a classification model (or
"classifier") on a test data set for which the true values are known. The confusion matrix is between
test labels and prediction labels. Fig 12 illustrates the confusion matrix where TP is an acronym for
True Positive, FP for False Positive, FN for False Negative and TN for True Negative.
Figure 12: Confusion matrix between test labels and prediction labels
11
The accuracy achieved using the MLXG model is 94% as shown in figure 9. The accuracy is
calculated using the definition as True Positive (TP) when the cases are predicted to have a skin
disease; true Negative (TN) when the cases predicted not to have a skin disease. False Positive (FP)
when the cases predicted to have a skin disease but actually have different disease not the one
predicted; false Negative (FN) when the cases predicted not to have a skin disease but actually they
have the skin disease. Figures 13, 14 and 15 show the output of the image and the predicted output
from the machine learning algorithm and the actual label of the image is to be given.
Figure 13: Output using MLXG model
Figure 14: Output using MLXG model
12
Figure 15: Output using MLXG model
5. Conclusion and Future Scope
This article introduced an improved classification system for detecting human skin disorders using
techniques of machine learning and this paper presented focuses on exploring a unique approach to
human skin disease detection using the same. As far as we know, our approach has not been evaluated
in the disease-related studies selected for the detection of human skin diseases. This proposed method
achieves 94% accuracy and great performance with the combination of feature extractor VGG16
which is a part of deep learning and machine learning classifier XGBoost for classifying the
problems. The proposed system can correctly identify acne, acne scars, alopecia areata, vitiligo and
warts. The MLXG method employs a deep learning technique by learning features of an image
through every single pixel using a pre-trained VGG16 network. The classification is then performed
by running the XGBoost classifier on the basis of extracted feature vectors. We can reach the
conclusion that this technique can be potentially used for the classification of diagnostic images when
developing MLXG expert systems. An average of 94% accuracy was achieved in the detection of
human skin diseases.
This paper attempts a new study of the combination of VGG16 and XGBoost. The proposed model
is built using a limited set of images in the dataset. We can test this model on big datasets as it doesn’t
use more space and the time taken to give the output is faster than other models. The skin diseases
selected are unique and the prediction given by the model is 100% accurate and can be used by the
dermatologist. In modern times, the efficient, proper and right use of technology is required. This
project will help develop the technological infrastructure of our nation. It will also support the Digital
India Campaign which was launched by our honourable Prime Minister Mr. Narendra Damodardas
Modi.
6. References
[1] “SKIN DISEASE DETECTION USING COMPUTER VISION AND MACHINE LEARNING
TECHNIQUE.” https://ejmcm.com/article_2063.html (accessed Jan. 13, 2022).
[2] “Skin Disease detection based on different Segmentation Techniques | IEEE Conference
Publication | IEEE Xplore.” https://ieeexplore.ieee.org/document/8862403 (accessed Jan. 13,
2022).
[3] “IRJET- Skin Disease Identification using Image Processing and Machine Learning Techniques
by IRJET Journal - Issuu.” https://issuu.com/irjet/docs/irjet-v7i3265 (accessed Jan. 13, 2022).
13
[4] N. Alenezi, “A Method of Skin Disease Detection Using Image Processing and Machine
Learning,” Procedia Comput. Sci., vol. 163, pp. 85–92, Jan. 2019, doi:
10.1016/j.procs.2019.12.090.
[5] “Disease Recognition Method Based on Image Color Skin and Texture Features.”
https://www.hindawi.com/journals/cmmm/2018/8145713/ (accessed Jan. 13, 2022).
[6] “Deep convolutional neural network for face skin diseases identification | IEEE Conference
Publication | IEEE Xplore.” https://ieeexplore.ieee.org/document/8940336 (accessed Jan. 13,
2022).
[7] K. Melbin and Y. J. V. Raj, “An Enhanced Model for Skin Disease Detection using Dragonfly
Optimization based Deep Neural Network,” in 2019 Third International conference on I-SMAC
(IoT in Social, Mobile, Analytics and Cloud) (I-SMAC), Dec. 2019, pp. 346–351. doi: 10.1109/I-
SMAC47947.2019.9032458.
[8] Ma. C. R. Navarro, E. Bustillos, and D. P. Y. Barfeh, “Skin Disease Analysis using Digital
Image processing,” in 2019 International Conference on Computational Intelligence and
Knowledge Economy (ICCIKE), Dec. 2019, pp. 311–316. doi:
10.1109/ICCIKE47802.2019.9004267.
[9] Y. Wang, J. Cai, D. C. Louie, H. Lui, T. K. Lee, and Z. Jane Wang, “Classifying Melanoma and
Seborrheic Keratosis Automatically with Polarization Speckle Imaging,” in 2019 IEEE Global
Conference on Signal and Information Processing (GlobalSIP), Nov. 2019, pp. 1–4. doi:
10.1109/GlobalSIP45357.2019.8969331.
[10] T. A. Rimi, N. Sultana, and Md. F. Ahmed Foysal, “Derm-NN: Skin Diseases Detection Using
Convolutional Neural Network,” in 2020 4th International Conference on Intelligent Computing
and Control Systems (ICICCS), May 2020, pp. 1205–1209. doi:
10.1109/ICICCS48265.2020.9120925.
[11] A. Haddad and S. A. Hameed, “Image Analysis Model For Skin Disease Detection: Framework,”
in 2018 7th International Conference on Computer and Communication Engineering (ICCCE),
Sep. 2018, pp. 1–4. doi: 10.1109/ICCCE.2018.8539270.
[12] V. Bevilacqua, A. Brunetti, A. Guerriero, G. F. Trotta, M. Telegrafo, and M. Moschetta, “A
performance comparison between shallow and deeper neural networks supervised classification
of tomosynthesis breast lesions images,” Cogn. Syst. Res., vol. 53, pp. 3–19, Jan. 2019, doi:
10.1016/j.cogsys.2018.04.011.
[13] X. Y. Liew, N. Hameed, and J. Clos, “An investigation of XGBoost-based algorithm for breast
cancer classification,” Mach. Learn. Appl., vol. 6, p. 100154, Dec. 2021, doi:
10.1016/j.mlwa.2021.100154.
[14] “VGGNet-16 Architecture: A Complete Guide | Kaggle.”
https://www.kaggle.com/blurredmachine/vggnet-16-architecture-a-complete-guide (accessed Jan.
13, 2022).
[15] J. Parashar, Sumiti, and M. Rai, “Breast cancer images classification by clustering of ROI and
mapping of features by CNN with XGBOOST learning,” Mater. Today Proc., Nov. 2020, doi:
10.1016/j.matpr.2020.09.650.
[16] “XGBoost | Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining.” https://dl.acm.org/doi/10.1145/2939672.2939785 (accessed Jan.
13, 2022).
[17] “Simple guide to confusion matrix terminology.” https://www.dataschool.io/simple-guide-to-
confusion-matrix-terminology/ (accessed Jan. 13, 2022).
[18] C. Flohr and R. Hay, “Putting the burden of skin diseases on the global map,” Br. J. Dermatol.,
vol. 184, no. 2, pp. 189–190, 2021, doi: 10.1111/bjd.19704.
[19] L. Tizek, M. c. Schielein, F. Seifert, T. Biedermann, A. Böhner, and A. Zink, “Skin diseases are
more common than we think: screening results of an unreferred population at the Munich
Oktoberfest,” J. Eur. Acad. Dermatol. Venereol., vol. 33, no. 7, pp. 1421–1428, 2019, doi:
10.1111/jdv.15494.
14