Image Recognition of Pigmented Skin Diseases Based on Deep

Learning

Xinyu Zhao

College of Electronic and Information Engineering, South China University of Technology,

Guangzhou, China

Keywords: Deep Learning, Pigmented Skin Diseases, Image Recognition.

Abstract: One of the most common skin conditions is pigmentary skin disease. It is also challenging to differentiate

between the lesions of various pigmentary skin diseases with the unaided eye due to their striking similarity.

The paper wishes to investigate whether deep learning image recognition can resolve this issue because deep

learning technology has advanced significantly in recent years and has shown promise in a number of

domains. In order to help the investigation, the paper modified the weights of three pigmented skin illnesses

that have similar clinical features to help two deep learning models that paper used to identify to gain higher

accuracy. The findings demonstrate that deep learning can effectively identify many forms of pigmented skin

illnesses and is very helpful in the recognition of skin diseases.

In subsequent research, the paper will attempt

to use deep learning to determine the lesion's stage, which will be extremely beneficial for diagnosing

pigmented skin conditions.

1 INTRODUCTION

The rapid evolution of computer vision and deep

learning technologies has brought significant

advancements to medical image analysis, particularly

in the detection and evaluation of pigmented skin

diseases. Pigmented skin conditions, which are

characterized by abnormalities in skin pigmentation,

represent a common category of dermatological

disorders. (Cai, 2023) The early detection and precise

staging of these conditions are critical for effective

clinical intervention and prognosis assessment.

Traditionally, the diagnosis of skin lesions has relied

heavily on the expertise of dermatologists, because it

is challenging to recognize and categorize skin

lesions due to the wide range of pigmented skin

lesions, the high degree of resemblance between

distinct classes, and the significant differences within

the same class. (Chen, 2014) This conventional

approach, while valuable, is often subjective and can

suffer from variability in diagnostic accuracy due to

differences in clinical experience and judgment.

In recent years, the integration of deep learning

techniques, particularly Convolutional Neural

Networks (CNNs), has introduced a transformative

https://orcid.org/0009-0009-1026-3190

shift in the analysis of medical images. CNNs have

demonstrated exceptional performance in tasks

related to image classification and object detection,

surpassing traditional methods in accuracy and

efficiency (Chu, 2024). These advancements present

promising solutions for the automatic identification

and staging of pigmented skin lesions, offering

potential improvements in diagnostic consistency and

operational efficiency.

This paper explores the application of deep

learning technologies to the recognition and staging

of pigmented skin lesions. The paper will examine the

current state of deep learning models used in skin

lesion image analysis, highlighting their applications,

the challenges they face, and their future prospects.

The discussion will focus on several key areas: the

mainstream deep learning methods employed,

including various CNN architectures; the processes

involved in constructing and processing datasets; and

the strategies for training and optimizing models.

Furthermore, the paper will consider the practical

implications of these technologies, evaluating how

they can be integrated into clinical practice to

enhance diagnostic capabilities and improve patient

outcomes.

Zhao and X.

Image Recognition of Pigmented Skin Diseases Based on Deep Learning.

DOI: 10.5220/0013512900004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 221-226

ISBN: 978-989-758-754-2

221

By providing a thorough examination of these

aspects, this paper aims to offer insights into the

current advancements in deep learning for skin lesion

analysis and to outline the potential for future

developments in this rapidly evolving field.

2 PIGMENTED SKIN LESION

DETECTION

Pigmented skin lesions (PSLs), a common type of

skin problem, are known to show more skin coloring,

mostly happening because melanocytes start to

multiply more. Proper detection and labeling of these

lesions are important as they come in both harmless

types, like melanocytic nevi (or just nevus), and

harmful ones like melanoma. Especially melanoma, a

very serious kind of skin cancer, is usually the cause

of many deaths related to skin cancer. It becomes

quite important to detect PSLs early on and categorize

them correctly for better treatment options and

improved health results for patients.

2.1 Types of hyperpigmented lesions

Pigmented skin lesions can mainly be broken down

into either benign or malignant forms. The usual

examples include the most common ones.

Melanocytic nevus: It is where melanocytes

gather in a benign way, showing up as dark spots on

the skin. Most of the time, they do not cause harm,

but there can be rare occasions where they turn into

melanoma.

Melanoma: A very dangerous type of skin cancer

that comes from melanocytes. Melanoma has a high

level of aggression, spreading to other body parts

quickly, so detecting it early is very important. (Han,

2018)

Sunspot: often referred to as liver spot or even

senile spot, it is mostly considered as a harmless

pigmented area that is the result of prolonged

exposure to the sun. These spots do not develop into

cancer, though sometimes they can be mistaken for

dangerous lesions of a malignant kind.

Seborrheic keratosis: This is identified as a non-

cancerous growth that is wart-like in its appearance,

which can form on different parts of the body.

Though these growths pose no harm, their visual

similarity to melanomas creates certain issues during

diagnosis.

Abnormal nevus: A kind of nevus that has

irregular characteristics and may suggest early signs

of melanoma, and regular check-ups become

necessary to watch for possible changes in these skin

areas.

2.2 Limitations of traditional

recognition methods

The common ways of identifying pigmented lesions

on skin mainly depend on doctors visually inspecting,

sometimes with help from dermoscopy, which is a

tool used to magnify the skin surface for better

viewing without causing harm. (Niu, 2024) Even

though these approaches can work well, they still face

certain limits and are not always fully sufficient.

Subjectivity: The evaluation of visuals is very

subjective, depending much on the experience and

expertise of dermatologists, which introduces

variations and inconsistencies in diagnoses made.

Limited accuracy: Even with dermoscopy being

used, some lesions remain difficult to tell apart,

particularly when they are in early development or

show unusual characteristics, adding to the difficulty

in making precise distinctions.

Time being consumed: Examining several lesions

manually, particularly when a patient has many

pigmented spots, takes much time and can be not

practical in certain clinical environments due to how

long it might take to complete.

Variability among observers: Different

dermatologists might see one lesion in various ways,

which creates inconsistency in diagnoses and advice

for treatments provided across different patients and

practitioners.

2.3 Image Classification Methods

Through Deep Learning

The traditional methods have limitations that

deeplearning methods aim to overcome, especially by

using CNNS more and more for the purpose of

classifying skin lesions with pigments.

2.3.1 Convolutional Neural Network (CNN)

Basic Idea

CNNS represent a kind of deep learning models

designed to handle structured grid-like data, for

example, images. A CNN typically consists of several

layers, which might include convolutional layers,

pooling layers, and layers that are fully connected.

Convolutional layers: In these layers, convolution

operations get applied to the input, with filters

detecting various local patterns like edges or shapes,

possibly textures. What results from this is a feature

map, which highlights such patterns but does so

without strict detail. (Dong, 2017)

Pooling layer: Pooling reduces the size of the

feature map by summarizing certain features present

DAML 2024 - International Conference on Data Analysis and Machine Learning

222

in regions, which has the function of lowering

computational complexity and also reducing

overfitting risks, although the exact effect is not

always completely clear.

Fully connected layers: Each neuron in one layer

gets connected with every neuron that belongs to the

next layer, allowing predictions to be formed based

on features coming from the convolutional and

pooling layers, though these predictions can be

influenced by many factors at the same time, some of

which could change.

2.3.2 Common Deep Learning models

There have been many different deep learning models

that have managed to be applied for the classification

of skin lesions that are pigmented, which shows the

versatility of these models in handling tasks related to

this field.

ResNet: ResNet brings up the idea of residual

learning, which makes networks able to go much

deeper by reducing issues with vanishing gradients. It

has been said to perform better in tasks like image

classification, including images from medical fields,

showing better results.

EfficientNet: EfficientNet represents a group of

models that, through scaling dimensions like depth,

width, and resolution in a structured manner,

improves performance. These models manage to

reach higher accuracy levels while using fewer

parameters than what is commonly seen in older

models. The approach provides better results not just

by increasing one aspect, but by adjusting several

dimensions together. Compared to more traditional

models, which may not consider such structured

scaling, EfficientNet shows advantages in both

efficiency and accuracy, making it stand apart. (Petra,

2024) However, exact improvements can vary

depending on implementation.

3 EXPERIMENTAL PRINCIPLE

3.1 Experimental procedure

As shown in Figure 1. Images and the associated

labels (melanoma, chromatoma, seborrheic keratosis)

make up the training set. The paper applies a Softmax

loss function and modify the weights according to the

mistake.

The images are routed to the convolutional,

pooling, fully connected (FC), and Softmax output

layers of ResNet101.

The model's performance is verified using the

images and labels that make up the verification set.

Images pass through the ResNet101 layers in a

manner akin to that of training.

Following the output's progression through the

ResNet101 layers, a test set is used to optimize and

validate it. To increase classification accuracy, the

optimization model modifies the parameters in

response to the output.

Figure 1:

Flow chart of Deep learning

(Photo/Picture credit : Original)

3.2 The Role of Deep Learning in Stage

Detection

Deep learning models are becoming more significant

in how pigmented skin lesions are automatically

detected and staged, helping with:

3.2.1 Feature Extraction and Discrimination

CNNs, which are used in many cases, can find and get

important features out of dermoscopic images. These

features, which are helpful in separating the stages of

melanoma, do not only include how the lesion looks,

but also other patterns that might exist, though these

patterns may not be easily noticed by human

observers. (Wang, 2024) These patterns can

sometimes be less obvious and need more attention to

be seen clearly.

3.2.2 Semantic and Instance Segmentation

Deep learning techniques, such as the more advanced

ones including semantic segmentation and also

instance segmentation, give the possibility of a

clearer outline of lesion boundaries, as well as finding

different areas inside a lesion that might match up

with various pathological traits. Semantic

segmentation is used to assign a label to each pixel in

Image Recognition of Pigmented Skin Diseases Based on Deep Learning

223

the image, which can help with recognizing certain

areas like tumor tissue in contrast with healthy parts.

But instance segmentation goes even further to

separate different objects or lesions found in one

image, though they might still fall under the same

general category.

3.3 Current Research and Applications

Recent research shows that deep learning models

have possibilities in automatically classifying and

staging pigmented skin lesions. For instance, some

researchers have created CNN-based models, which

can sometimes reach accuracy levels close to or even

higher than experienced dermatologists when it

comes to detecting melanoma. Additionally, models

that were trained on big annotated datasets, such as

the ISIC dataset (International Skin Imaging

Collaboration), are now used to automate the staging

process for melanoma. (Niu, 2024) This offers a

useful tool to help clinicians with decisions when

making diagnoses.

4 EXPERIMENT AND RESULT

4.1 Construction and Choosing of

Datasets

The deep learning model's effectiveness when it

comes to detecting and staging pigmented skin

lesions relies a lot on the type of dataset and how

varied it is, especially for training purposes.

Important points to think about include:

Data Collection: High-resolution images must be

gathered from different groups of people and medical

settings. These images are needed for creating

reliable models. The images should cover different

phases of illness and show both frequent and unusual

types of lesions, which is important to ensure the data

is comprehensive enough for the purpose of training.

Annotation: It is necessary to annotate datasets with

accuracy. Usually, experts like dermatologists are

responsible for labeling images, and they mark them

with details like lesion categories, disease stages, and

important clinical attributes. In some situations, the

annotations may even provide pixel-wise labeling for

tasks related to segmentation.

The data set used in this study was HAM10000

skin image dataset on hyperai(shows on Figure 2),

which collected 10000 skin lesions images from

different populations. The cases included several

representative types of hyperpigmented lesions,

mainly melanoma, nevus and seborrheic keratosis.

HyperAI (hyper.ai) artificial intelligence and high

performance computing community aims to help

developers and enthusiasts of data science and

artificial intelligence industry learn, understand and

practical by providing multiple services such as

accelerated download of data sets, online tutorial

demonstrations, in-depth interpretation of papers, and

integration of top conference calendar.

Figure 2: Dataset from Hyperai(Photo/Picture credit :

Original)

4.2 Model Training and Performance

Evaluation

It is very important to do model training and

performance evaluation correctly if you want deep

learning models to be used in clinical practice.

4.2.1 Choosing and Preparing the Training

data

After removing the unidentifiable bad graphs in the

dataset, the paper divided the dataset into three

independent datasets that did not cross each other to

ensure the generalization ability of the model on

unknown data. There were 6000 images in the

training set, 1800 images in the validation set, and

450 images in the test set.

The distribution of three types of tumor images

on the three datasets is shown in table 1.

DAML 2024 - International Conference on Data Analysis and Machine Learning

224

Table 1: Dataset distribution

Dataset Melan-

oma

Chrom-

atoma

Seborrh

-eic

keratosi

-s

Total

Training

set

1800 2400 1800 6000

Verificati

-on set

540 720 540 1800

Test set 145 180 135 450

4.2.2 Model Evaluation Metrics

Metrics commonly used for evaluating how models

perform include:

Accuracy: It is the ratio of correct predictions

compared to the total number of predictions that the

model has made overall.

Sensitivity (Recall): This measures how well the

model can find positive instances, like when it

identifies malignant lesions correctly.

Precision: Precision is a metric that is used in

statistical classification and information retrieval that

represents the proportion of correctly extracted

samples to total extracted samples. (Tian. 2024)

Recall, which is the ratio of the number of extracted

samples to the total number of samples, is a related

idea.

F1 Score: Precision and recall come together to

make a harmonic mean, and that gives a measure

which balances how the model performs on both

sides, providing an understanding of overall

performance.

4.3 Experimental result

During the training phase, the model's prediction

performance was assessed in real time using the

validation set, which allowed for parameter

adjustments and overall model optimization. The

final model for predicting the test set data was

determined by using the training result from the round

that had the highest accuracy of the validation set.

There were 100 training epochs in this model.

ResNet101 was used as the training model in this

experiment in order to confirm the experiment's

efficacy. By establishing category weights, the

efficacy of deep learning in recognizing photos of

pigmented skin diseases was assessed. This

experiment is trained on a Tesla V100S-PCI-32GB

GPU and is based on the Aliyun computing platform

to guarantee experiment efficiency.

It can be seen from Figure 3 that the accuracy of

diagnosing pigmentary skin illnesses improved to

65% after many training sessions, indicating that deep

learning is a viable method for doing so. The figure 4

compares the particular recognition results of each

category when the two training procedures are used

with this model.

Figure 3:

Curve of average accuracy(Photo/Picture credit :

Original)

Figure 4: Accuracy of 3 kinds of skin diseases

(Photo/Picture credit : Original)

Deep learning had generally good accuracy in

identifying chromopoma and seborrheic keratosis,

but less desirable results when it came to melanoma.

0.2

0.4

0.6

0.8

0 20406080100120

epochs

accuracy

Image Recognition of Pigmented Skin Diseases Based on Deep Learning

225

5 CONCLUSION

One of the most crucial responsibilities in

dermatology is the diagnosis of pigmented skin

diseases, which has a big influence on patient

outcomes. While useful, traditional diagnostic

techniques have drawbacks that deep-learning

models might be able to solve. These models are able

to correctly classify partly pigmented lesions through

the use of sophisticated techniques including

segmentation algorithms and CNN. Many obstacles

still need to be overcome, particularly in the areas of

data heterogeneity, model applicability, and

distinguishing between different kinds of pigmented

dermatoses. It will be up to continued study and

creative problem-solving to resolve these obstacles

before deep learning in dermatology can realize its

full potential.

The paper employed two distinct deep learning

models to identify three common pigmented skin

illnesses from images. By varying the weights of the

models, the paper was able to increase the recognition

accuracy. The encouraging outcomes demonstrated

that deep learning may be used to diagnose skin

illnesses, with up to 80% accuracy being able to

distinguish between chromoblastoma and seborrheic

keratosis. The paper will attempt to use deep learning

to distinguish between various phases of pigmented

skin lesions and to increase the accuracy of

diagnosing skin disorders like melanoma, which are

challenging to diagnose.

REFERENCES

Cai, Q. Y., Ma, Q., Wang, C. K. 2023 Application of

deep learning in skin OCT medical images.

Chinese Journal of Laser Biology,32(3): 193-199.

Chen, X. C. 2014. Deep Learning Algorithm and

application Research based on Convolutional neural

networks. (Doctoral dissertation, Zhejiang Gongshang

University)、

Chu, J. L., Fang, C., Wan, N., Wang, W. D., Wang,

X., Zhang, J. 2024. Application of deep learning

in classification and recognition of pneumonia CT

images. Fujian computers (07), 33-36. Doi:

10.16707 / j.carol carroll nki FJPC. 2024.07.006.

Dong, J., Jin, L. P., Zhou, F. Y. 2017. A review of

convolutional neural networks. Journal of

Computer Science, 40(6), 1229-1251.

Han, Z. Y., He, X. Y., Wei, Z. Z. 2018 Recognition

and classification of pigmented skin diseases

based on deep convolutional neural networks.

Journal of Computer Applications,38(11): 3236-

3240.

Nadia, S., Razieh, P., Sajad, S., & José L. H. 2024.

Identification of Armyworm-Infected Leaves in

Corn by Image Processing and Deep

Learning.Acta Technologica Agriculturae(2),92-

100.

Niu, H., Wang, P. Y., Yin, G. W. 2024.

Clinicopathological analysis of 15 cases of basal

cell carcinoma of the skin. Oncology

Fundamentals and Clinic (02),200-202.

Petra, R., Dorijan, R. & Goran, M. 2024. Image-

Based Leaf Disease Recognition Using Transfer

Deep Learning with a Novel Versatile

Optimization Module. Big Data and Cognitive

Computing(6),52-.

Tian, S. W., Shi, X. W., Shu, C., Yu, L. 2024.

MSMA: A multi-stage and multi-attention

algorithm for the classification of multimodal skin

lesions, Biomedical Signal Processing and

Control, Volume 93,106180, ISSN 1746-8094.

Wang, H., Wang, T. T. Wu, J. H., Yang, L. 2024.

Diagnosis of hand arthritis based on deep learning

algorithm from X-ray images. Modern Medicine

(07),1043-1049.

DAML 2024 - International Conference on Data Analysis and Machine Learning

226