A Web-Based Skin Disease Diagnosis Using Convolutional Neural Networks

Published Online November 2019 in MECS DOI: 10.5815/ijitcs.2019.11.06

The human skin plays a huge part in a person's physical appearance, and it is the biggest organ of the human body. The human-skin offers protection against fungal infection, bacteria, allergy, viruses and controls the temperature of the body situations that change the texture of the skin or damage the skin can produce symptoms like swelling, burning, redness, and itching [1]. Antipathies, genetic structure, irritants, and particular diseases and immune system associated complications can produce hives, dermatitis, and other skin problems. Many of the skin diseases, such as alopecia, eczema, acne, ringworm, also affect a person’s look. Skin diseases are widespread these days; some of them are simple and easy to recover from; others are very harmful and might be incurable, and many of these diseases are very dangerous, mainly if not treated in the early stages.

In literature, dermatological diseases are reported to be the most widely spread disease [2]. A survey by Hogewoning et al. [3] revealed that the total frequency of pupils with some skin disease was 34.6% and 42.0% in two (2) Ghanaian studies, out of 4,839 pupils surveyed. Again, in [4] reports, out of 529 participants surveyed, 700 discrete skin diagnoses were made [4]. Therefore, detection of skin disease at its early stages is paramount to its spreading.

On the other hand, skin disease diagnosis is seen to be complicated, mainly when two or more diseases portray same or similar symptoms, hence requires a dermatologist with vast experience of skin diseases [2,4]. Nevertheless, the development in technology and machine learning have changed all aspects of one’s dayto-day life, including the medical field [5,6]. Many therapeutic systems have been developed with the help of artificial intelligence (AI) and technological advancement to help both doctors and patients in diverse ways, starting from Out Patient Department (OPD), consultation to the operating theatre or operating room (OR). Thus, the introduction of artificial intelligence into the health industries has brought tremendous improvement in the diagnoses of skin disease and other illness [7].

However, in Ghana, most dermatologists still use a variety of manual visual clues such as colour, scaling, and arrangement of the lesions, the body site distribution, among others. Nonetheless, when these individual components are analysed separately, the recognition of the disease can be quite complex, thus requiring a high level of experience. Human diagnosis is based on a subjective judgment of the dermatologist, so it is hardly reproducible, unlike computer-aided diagnostic systems, which are more realistic and reliable.

To reduce diagnosis time and provide quick health service, some researchers in recent years proposed skin disease detection system with the ability to detect skin disease like impetigo, eczema, melanoma and acne using machine learning [8–10] . On the other hand, these skin diseases are not prevalent in Ghana, as indicated in [4].

Furthermore, Ghana currently has only one dermatology-training centre at the Korle Bu Teaching Hospital (KBTH), with only four (4) dermatologists three

• (3)    on full-time and one (1) part-time, and three (3) trainees. In reality, the total number of dermatologists serving the whole people of Ghana is lesser than 25 [4]. On average, a patient with skin disease spent not less than two hours in a medical centre.

Finally, with a population 30,030,189 as of May 2019 [11], it implies that every dermatologist in Ghana is to 1,201,207.56 patients.

In an attempt to reduce issues mentioned above, the current study seeks to develop a smart web-based skindisease detection system (medilab-plus) for faster and reliable early detection of atopic dermatitis, acne vulgaris, and scabies, using the convolutional neural network (CNN).

The development of the proposed system will offer foreknowledge, quick and faster diagnosis system to users through the internet. Again, serve as the first skin diseases system built and tested with sample data from Ghana.

The remaining section of this study is categorised as follows: Section 2 present review of common skin diseases in Ghana, the application of machine learning in disease diagnosis system, and related studies. Section 3 covers the methods, tools and evaluation metrics adopted for the current study. Section 4 presents the outcome and discussion of the study. Finally, Section 5 concludes the study and the direction for future studies.

• II.    Literature Review

This section gives a brief discusses of common skindisease in Ghana, the application of machine learning in disease diagnosis system, and related studies.

• A.    Skin diseases in Ghana

There are numerous types of skin diseases identified in Ghana. However, table 1 shows the common skin diseases in Ghana and their prevalence among males and females, as presented in [4].

Table 1. Common skin diseases in Ghana

Overall ( N=700)	N (%)	Males ( N=302)	N (%)	Females ( N=396)	N (%)
Atopic dermatitis	59 (8.4)	Atopic dermatitis	24 (7.9)	Atopic dermatitis	35 (8.8)
Acne vulgaris	37 (5.3)	Scabies	17 (5.6)	Pityriasis rosea	21 (5.3)
Scabies	36 (5.1)	Warts	17 (5.6)	Lichen planus	21 (5.3)
Irritant contact dermatitis	33 (4.7)	Acne vulgaris	16 (5.3)	Acne vulgaris	20 (5.1)
Lichen planus	26 (3.7)	Irritant contact dermatitis	14 (4.6)	Scabies	19 (4.8)
Seborrhoeic dermatitis	25 (3.6)	Seborrhoeic dermatitis	11 (3.6)	Irritant contact dermatitis	19 (4.8)
Warts	23 (3.3)	Tinea pedis	10 (3.3)	Vitiligo	14 (3.5)
Vitiligo	22 (3.1)	Pityriasis versicolor	9 (3.0)	Papular urticaria	13 (3.3)
Pityriasis versicolor	17 (2.4)	Chronic urticaria	9 (3.0)	Seborrhoeic dermatitis	13 (3.3)

• B.    Methods of identifying skin diseases

Readily visible changes of the skin surface have been recognised since the genesis of history, with some treatable, and some not. In developing countries, overcrowding and poor hygiene are responsible for spreading of skin diseases. One of the known initially sources detailing skin diseases is the Ebers Papyrus, a medical paper from antique Egypt dating to around 1500 BC. It offers descriptions of the various skin diseases, including ulcers, rashes, and tumours, and prescribes surgery and ointments to treat the ailments [12]. There are two ways of detecting or diagnosing skin disease

The first method is the traditional method, also known as the conventional method in which skin diseases are detected based on unique colour space. Due to the mixing of chrominance and luminance data, RGB is not the right choice for detection. Although it avoids this problem, its actual detection effect is still unstable and susceptible to some environmental influences [13,14]. The specific positioning of the affected area is necessary to detect the type of skin disease.

The second method is the technological method, with the emergence of machine learning, diagnosing of skin disease has become easy for most dermatologists. Computer Vision (CV), Machine-Learning, and Artificial Intelligence are the approach introduce on clinically evaluated histopathological attributes to identify the condition accurately. Firstly, the image is pre-processed, followed by feature extraction. The second stage involves the use of machine-learning algorithms to classify conditions based on the histopathological attributes observed on the analysing of the skin.

A CV is an interdisciplinary field that concerns with how computers can be made to gain a high-level understanding from digital images and videos. From the engineering perspective, it seeks to automate tasks that the human visual system can do. Sub-domains of CV include scene reconstruction, event detection, video tracking, object recognition, object pose estimation, learning, indexing, motion estimation, and image restoration [1,6].

• C.    Machine Learning (ML)

ML is a subcategory of artificial intelligence, which, uses statistical and computational tools to offer humanlike abilities to computers [17,18]. Thus, ML offers automation and enhancement of the learning process of machines based on their experiences without being programmed (no human assistance) [19]. ML Techniques can be grouped into three main categories, namely; supervised, unsupervised, and reinforcement learning [20,21]. ML algorithms such as decision trees (DT), artificial neural network (ANN) support vector machines (SVM), Naïve Bayes and AdaBoost has been applied in the various disease-diagnosis system [1,8,22] . Fig. 1 shows a comparison of accuracy between some machine learning algorithms in disease detection from 2009 to 2015.

Fig.1. Accuracy of machine learning algorithms in disease detection (source: [22]).

• D.    Related works

A skin disease diagnosis system was proposed in [8] , where a user uploads an image of the affected area of the skin into an online system and receive treatment or a piece of advice in a short time . An empirical result of their system offered an accuracy of 95% for Impetigo, 85% for Eczema, and 85% for Melanoma.

An Android app (skinvision) for the diagnosis of melanoma skin disease was proposed in [15]. The accuracy level of the skinvision app was 81%. Users of the app take an image of the disease spot with a phone camera and upload it into the app, and a verdict is given within 30 seconds as low, medium, or high risk. A Melanoma skin cancer detection model was proposed in [9], using a support vector machine.

Similarly, Rathod et al. [2] proposed an application for detecting five (5) different skin disease using convolutional neural networks. The proposed system attained an accuracy of 70%. However, the authors concluded that using a higher dimensional dataset can increase the accuracy above 90% [2].

Table 2 shows a summary of related studies. Most previous studies were based on the classification of Melanoma and Eczema [8-9] and [16]. Only a few studies [10] were carried out on acne detection. However, these diseases are not prevalent in Ghana as discussed in section 2.1.

Table 2 . Compilation of Related Methods