A Streamlined Lesion Segmentation Method Using Deep Learning and

Image Processing for a Further Melanoma Diagnosis

Jinen Daghrir

1 a

, Wafa Mbarki

, Lotﬁ Tlig

, Moez Bouchouicha

, Noureddine Litaiem

Faten Zeglaoui

and Mounir Sayadi

Universit ´e de Tunis, ENSIT, Laboratory SIME, Tunisia

Universit ´e de Toulon, Aix Marseille Univ, CNRS, LIS, Marseille, France

University of Tunis El-Manar, Faculty of Medicine of Tunis, Department of Dermatology,

Charles Nicolle Hospital, Tunis, Tunisia

Keywords:

Deep Learning, u-Net, Image Processing, Skin Lesion Segmentation, Majority Voting, Medical Imaging.

Abstract:

Over the past two decades, the world has known a signiﬁcant number of deaths from cancer. More speciﬁcally,

melanoma which is considered as the deadliest form of skin cancer causes a remarkable percentage of all can-

cer deaths. Therefore, the health and disease management community has exceedingly invested in creating

automated systems to help doctors better analyze such diseases. Correspondingly, we were interested in creat-

ing an automatic lesion detection task for further melanoma diagnosis. The lesion segmentation is considered

to be a critical step in a pattern recognition system. Our proposed segmentation technique consists of ﬁnding

lesions’ masks using a baseline, edge-based, and more sophisticated and state-of-the-art method: thresholding

using Otsu’s technique, morphological snakes, and a fully CNN (Convolutional Neural Network) model based

on the U-net architecture, respectively. These methods are commonly used when dealing with skin lesion seg-

mentation, and each one of them has its advantages and drawbacks. The U-net architecture is improved by the

use of the pre-trained encoder ResNet-50 on the ImageNet dataset. A majority voting is applied to generate

the ﬁnal segmentation map using these three methods. The experiments were conducted using a benchmark

dataset and showed promising results compared to using these methods separately, the majority voting of the

three methods can signiﬁcantly improve the segmentation task by reﬁning the borders of the masks issued by

the Deep learning model.

1 INTRODUCTION

Skin cancer is considered one of the most frequent

types of cancer. Melanoma which is a particular

form of skin cancer, is the less common type but it

is considered as the most malignant one. This fa-

tal skin cancer can quickly spread to other parts of

the body causing about 60 000 cancer deaths in 2018

as reported in (Khazaei et al., 2019), it is consid-

ered as 0.7% of all cancer deaths. Statistics exam-

ining the incidence rate from 1973 to 2009 show a

rise in the number of cases (Heinzerling and Eigentler,

2021), which is particularly worrying in such a dis-

ease. Some types of skin lesions can be more dan-

gerous than others, due to their ability to spread to

other sites. About 5 to 15% of people having con-

genital nevi are more risked to develop melanoma.

https://orcid.org/0000-0002-1300-8939

About one-third of cases will be certainly affected

in their brains and can be associated with malforma-

tion of the central nervous system (Heinzerling and

Eigentler, 2021). The incidence and mortality rates

of melanoma differ from one country to another due

to the variety of ethnic and racial groups(Schadendorf

et al., 2018). The well-known deﬁnite risk factors for

melanoma include ultraviolet (UV) radiation, life in

low geographic latitudes, high alcohol consumption,

consuming fatty foods, the presence of melanocytic

or dysplastic naevi, a family or personal history of

melanoma, phenotypic characteristics including fair

hair, eye, and skin colors (Khazaei et al., 2019). In-

terestingly, it has been proved that the development

index (HDI) is quite related to the number of mortal-

ity of melanoma. The more the HDI index increases,

the more people have access to health services and

early detection of diseases, and the less mortality will

occur(Khazaei et al., 2019). UV protection is recom-

368

Daghrir, J., Mbarki, W., Tlig, L., Bouchouicha, M., Litaiem, N., Zeglaoui, F. and Sayadi, M.

A Streamlined Lesion Segmentation Method Using Deep Learning and Image Processing for a Further Melanoma Diagnosis.

DOI: 10.5220/0013472400003938

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2025), pages 368-376

ISBN: 978-989-758-743-6; ISSN: 2184-4984

mended since it is proved that using protection will

lead to a decrease in the incidence rate. However,

the use of sunscreen does not provide enough pro-

tection against the development of melanocytic nevi

consequently the high risk of developing melanoma

(Heinzerling and Eigentler, 2021). One typical solu-

tion to decrease the annual number of deaths caused

by melanoma is diagnosed it in an early stage, where

the relative survival rate will be critical (Schaden-

dorf et al., 2018). Self-examination is so vital in this

case. Thus, people should examine their skin head

to toe regularly, looking for any lesions that might be

turned into melanoma. Self-exams can help people

identify skin cancers in an early stage when the odds

of curing them are completely high. Yet, physicians

encourage people to routinely do self-examination if

they notice any suspicious-looking lesions. ”when in

doubt, cut it out”, says Darrick Antell, MD, a board-

certiﬁed plastic and reconstructive surgeon practicing

in New York (Daghrir et al., 2020). Currently, the

health and disease management community has en-

tirely been interested in creating automated computer

systems for skin lesion inspection and characteriza-

tion. Thus, it becomes vital to use supportive imaging

to improve the diagnosis of this life-threatening skin

cancer. Conventionally, these systems use some vi-

sual cues to characterize the lesions as malignant or

not. One typical used clue is the ABCD signs. This

feature can differentiate malignant melanoma from

a benign skin lesion based on ﬁve deﬁned charac-

teristics, namely asymmetry (A), border irregularity

(B), color variability (C), and diameter greater than 6

mm (D)(Maglogiannis and Doukas, 2009). To extract

information about the border, lesion segmentation

should be performed. This task is considered to be

a critical step in the whole process of melanoma de-

tection, it concerns the isolation of pathological skin

lesions from the surrounding healthy skin. The seg-

mentation step is usually done by experts, who try to

ﬁnd the ROI (Region Of Interest). This method is ac-

curate since the experts apply their knowledge and ex-

periences to the segmentation process. Thought, this

process will be time-consuming and exigent, due to

the variability and reliability of the expert behavioral

observation. One straightforward method is to use

fully automatic systems where the segmentation pro-

cess is done using computer vision techniques without

the need for human intervention.

In medical image processing, ROI segmentation is

deﬁned as a semantic segmentation which is a pixel-

level recognition problem. It refers to the process of

assigning each pixel in the image to a particular label.

Particularly, in the case of skin lesion segmentation,

healthy skin is assigned a background label. How-

ever, the inspected lesion is assigned a foreground la-

bel. The result of this process is a segmentation map,

where the background and foreground regions are as-

signed with two different class labels. Commonly,

many classical image processing techniques are pro-

posed for Medical image segmentation (Mbarki et al.,

2020; Sharma and Ray, 2006). These techniques rely

only on low-level pixel processing and mathematical

modeling to construct rule-based systems. Thresh-

olding has been widely used since it is simple and

not time-consuming. It consists of dividing an im-

age into two regions: one corresponds to the ROI and

the other to the background. Thus, each pixel is as-

signed a class label if its intensity is greater than a

threshold. The threshold can be either globally or lo-

cally ﬁxed according to a small set of pixels’ inten-

sities. The threshold is deﬁned using statistical anal-

ysis or by optimizing a certain criterion. Thus, us-

ing only thresholding can often fail to produce accu-

rate results since it can generate disconnected objects.

However, Active contours or snakes can guarantee a

connected set of pixels for an ROI. These methods

are considered edge-based methods. An active con-

tour is formulated as an energy minimization prob-

lem, where the energy depends on the placement of

the curve surrounding the object to be segmented.

The initial contour is manually deﬁned and then it-

eratively deformed to minimize the energy function,

which is typically based on the intensity differences

between the object and the background pixels. Obvi-

ously, these methods are very sensitive to noise and

the initial curve placement. The large variability of

the shape of substructures and organs in medical im-

ages is one of the most challenging problems for the

segmentation task. Therefore, there is no universal

segmentation technique. Some techniques do not use

only color information of pixels, but also much more

domain knowledge(Litjens et al., 2017). In order to

deﬁne and implement this knowledge, a considerable

amount of expertise and time is required. Thus, com-

pletely automated systems are proposed to model the

domain knowledge by training some algorithms with

labeled pixels. Recently, a signiﬁcant trend in seg-

mentation tasks is the use of deep learning(DL)due

to its effectiveness and speediness. With the advance

of DL models, scientiﬁc papers presenting organ and

substructure segmentation have gradually increased in

the last few years(about 90 papers were published in

2017)(Litjens et al., 2017). The motivation for the

excessive use of DL is the ability to use one model

in many different domain applications and with dif-

ferent data resources, such as in(Alom et al., 2018),

authors have tested their proposed DL model on three

different benchmark datasets such as blood vessel in

A Streamlined Lesion Segmentation Method Using Deep Learning and Image Processing for a Further Melanoma Diagnosis

369

the retina, skin lesion, and lung lesion datasets. One

widely used architecture of DL is the CNN( Con-

volutional Neural Network) which ensures continued

good performance in many tasks, such as classiﬁca-

tion, segmentation, detection, decision-making, etc.

The number of papers published in 2017 applying

CNNs in medical image analysis is over 200 papers,

and most of them are dedicated to dealing with sub-

structure segmentation(Litjens et al., 2017). Excep-

tionally, for semantic segmentation, a fully convolu-

tional network is required, which contains two op-

posite units, encoder and decoder. The encoder is

used to deﬁne the high-level information with a low-

resolution spatial tensor, while the decoder contains

also convolution layers but coupled with up-sampling

instead of down-sampling layers that increase the size

of the spatial tensor accordingly recapturing the seg-

mented image. In(Ronneberger et al., 2015), the au-

thors have proposed the U-net architecture that takes

an input image and outputs a segmentation map that

has the same dimension as the input. The U-net was

ﬁrst used to detect cell boundaries in biomedical im-

ages. The contribution of such architecture is to use

skip connections, which connect opposite convolu-

tional layers in order to preserve low-level informa-

tion. Deﬁnitely, DL is efﬁciently and widely used

for many tasks, but the fact that it is time-consuming

and data-hungry makes it a little bit limited. The per-

formance of systems using DL will degrade if the

amount of input data is not big enough. In addi-

tion, the data can be only produced by an expert who

will label every input. Time and resources are re-

quired to properly do data labeling to build highly

accurate systems, but it is not a straightforward pro-

cess when it comes to huge datasets. For melanoma

diagnosis, the border-based features should be ex-

tracted after segmenting the skin lesion. Two main

automatic segmentation methods are presented in the

literature, region-based and contour-based methods.

Other hybrid methods are proposed which use both

color transformation and contour detection techniques

(Maglogiannis and Doukas, 2009). These methods

are also classiﬁed into two groups: Methods that use

classical image processing with analysis techniques

and those that involve Artiﬁcial Intelligence(AI) con-

cepts. The most usual and basic techniques that are

used for lesion segmentation, are thresholding and

region-growing techniques. These methods are clas-

siﬁed as region-based approaches since they involve

the differentiation of pixels’ intensities of both ma-

lignant and healthy skin regions. Another type of

method is contour-based which detects the region of

interest by ﬁnding its edge. However, with the ad-

vent of more powerful techniques, these techniques

seem to be very naive. For that, hybrid approaches

are proposed in the literature (Jamil et al., 2019)

that use both color transformation and edge detec-

tion, snakes or active contours are considered to be

the state-of-the-art of lesion segmentation before the

extensive use of machine learning(Maglogiannis and

Doukas, 2009). Other methods involving AI tech-

niques like fuzzy borders, K-means segmentation, or

Deep Learning (DL) techniques are considered nowa-

days as the state-of-the-art for most image segmenta-

tion problems (Jadhav et al., 2019). Melanoma de-

tection at an early stage can highly extend the life

expectancy, thus it becomes vital to use CAD to en-

sure a quick and accurate diagnosis of a huge number

of people. Due to the sensitivity of the lesion seg-

mentation process which plays an imperative role in

melanoma diagnosis. Many research papers are pro-

posed in the literature dealing with lesion segmenta-

tion, although research is still trending up sharply due

to a lot of challenges.

In this matter, we were interested in creating a

new method for lesion segmentation that combines

three different methods: thresholding using Otsu’s

technique, morphological Snakes, and a deep convo-

lutional neural network based on U-net. The afore-

mentioned techniques are used independently to gen-

erate three lesion masks. Then, using majority voting

a new lesion segmentation map is generated. Major-

ity voting deﬁnes the class of each pixel whether it

belongs to the lesion or to the background based on

the other decisions. This paper is organized as fol-

lows. In the following section, we extendedly intro-

duce our proposed method. All the conducted exper-

iments and the results evaluating the performance of

the proposed method are given in the third section. At

the end, we highlight and discuss the basic concept of

our proposed method.

2 PROPOSED METHOD

One critical process in melanoma diagnosis is lesion

segmentation. It consists of identifying the lesion’s

pixels among the others belonging to the healthy

skin. In this paper, a new method has been proposed

that combines three different methods using major-

ity voting. The fusion of the three results issued by

thresholding using Otsu’s technique, morphological

snakes, and CNN-based U-Net architecture, signiﬁ-

cantly ameliorates the segmentation mask by reﬁning

the lesion’s border.

ICT4AWE 2025 - 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

370

2.1 Thresholding Method

Thresholding has been widely used in skin cancer due

to its effectiveness and integrity. Researchers have

proposed many local adaptative and global threshold-

ing techniques, such as Kapur et al.’s Entropy tech-

nique(Armato et al., 2002), Abutaleb’s entropy tech-

nique(Ashwin et al., 2012), and Kittler and Illing-

worth’s minimum error technique(Bae et al., 2005).

However, Otsu’s algorithm has been extensively ap-

plied in many systems. Thresholding is done by ﬁnd-

ing one single threshold that classiﬁes pixels into two

classes without image preprocessing. The threshold is

determined by minimizing intra-class intensity vari-

ance σ

, which is deﬁned as the weighted sum of

class variances:

(t) = w

(t)σ

(t) + w

(t) (1)

and w

are the probabilities of the two classes

separated by a threshold t, and σ

and σ

are the

variances of the two classes. After computing the

probabilities of each class using the histogram, w

(0)

and σ

(0) are initially set. Using different values of

threshold t from 1 to the maximum pixel’s intensity,

(t) and σ

(t) will be computed to get the intra-

class intensity variance. The global threshold will be

the one that minimizes the σ

(t). This threshold is

used then to differentiate the foreground which is the

pathological skin lesion and the background. The seg-

mentation map is found by assigning the value 1 to all

the pixels having an intensity greater than the thresh-

old t and the value 0 to all the other pixels. Two steps

are performed to ﬁlter the unwanted objects. First, an

opening is used twice, the morphological opening is

an erosion followed by dilation using a disk as a struc-

turing element. The erosion of the image helps to re-

move the small objects. Consecutively, dilation helps

in ameliorating the background region by closing all

the gaps. For further ﬁltering, big regions having a

centroid that does not belong to 75% of the image area

will be eliminated. This process ensures that only the

skin lesion mask will stand. Fig.1 illustrates the result

of using thresholding and ﬁltering the segmentation

map.

2.2 Morphological Snake Method

Active contours, also known as Snakes are one of

the most used techniques in computer vision, more

speciﬁcally for image segmentation. Frequently,

these methods are edge-based which requires clean

boundaries indicting and the absence of artifacts and

noises. These problems were then handled by the

invention of more practical approaches such as the

Active Contours Without Edges (ACWE)(Marquez-

Neila et al., 2013). These approaches are known

as morphological snakes since they use morpholog-

ical operators over a binary array. The morpholog-

ical ACWE can accurately segment the object when

the pixels of the region of interest and those of the

background have different averages. In this work, the

Chan-Vese algorithm(Chan and Vese, 1999) which is

a form of the snake curve is adopted. It is an energy

minimization method that is widely used in medical

applications. This method does not require deﬁning

the contours to separate the heterogeneous objects and

it is sensitive to noise. This algorithm can also ac-

curately detect the objects and increase the conver-

gence speed when using a good initial curve place-

ment. Since this method can detect more than one

region of interest and obviously the skin lesion will

be placed in the center of the image, we have used the

initial curve presented in Fig.2. The result of using

the snakes is shown in Fig.2. The ﬁnal segmentation

mask is generated after postprocessing the output of

the Snake process. We have applied the morphologi-

cal opening to reﬁne the segmentation map.

2.3 U-Net Based Resnet50

The U-Net architecture is a fully convolutional neu-

ral network that was mainly developed for biomed-

ical image segmentation. It consists of two blocks:

encoder and decoder. CNNs are commonly used for

classiﬁcation tasks, the input image is fed to the net-

work and downsampled into a low-resolution tensor

containing high-level information which can then be

classiﬁed using a fully connected layer. However, to

generate a segmentation map that has the same size as

the input, the output of the encoder should pass by an

upsampling path( decoder). This makes the network

layout has a U shape.

The left side of the U follows the typical archi-

tecture of a convolutional network. It consists of the

repeated application of two 3× 3 convolutions, which

is followed by a normalization layer and a rectiﬁed

linear unit (ReLU) activation function. Dropout and

2 × 2 max-pooling layers are used to downsample the

feature map. At each downsampling step, the num-

ber of feature channels is increased. The purpose of

the encoder is to capture the high-level information of

the input image to do the segmentation process. To

prevent losing the low-level information, the decoder

should have connections with the encoder layers, that

is known as skip connections.

Many CNNs are previously trained for image clas-

siﬁcation tasks, that contain meaningful information

about how to perform predictions effectively. These

A Streamlined Lesion Segmentation Method Using Deep Learning and Image Processing for a Further Melanoma Diagnosis

371

Figure 1: Segmentation result of the thresholding process: (a) Original image, (b) segmentation map using Otsu’s algorithm,

and (c) segmentation map after ﬁltering.

Figure 2: Segmentation result of using morphological

snakes: (a) initial curve, (b) evolution of the curve, (c) ﬁ-

nal segmentation map, and (d) ﬁnal segmentation map after

applying morphological operations.

models can be re-used to vastly speed up the train-

ing process, this concept is called transfer learn-

ing. Further, the convolution layers of the pre-

trained models can be re-used in the encoder lay-

ers of the segmentation model. In this work, a

Resnet50 is adopted which was pre-trained on the

ImageNet dataset(Russakovsky et al., 2015). This

dataset is widely used when dealing with medical

imaging problems since it contains a repository of

annotated clinical (radiology and pathology) images.

Thus, the U-net model will use the weights of a

pre-trained ResNet50 model systematically having

knowledge about the disposed features.

CNN has known many improvements over time,

one common trend in research was to add more regu-

larization layers to overcome overﬁtting. AlexNet has

ﬁve convolutional layers, however, other architectures

have gone deeper to ameliorate their performance,

such as VGG and GoogleNet networks. Therefore,

increasing the convolutional layers does not regularly

improve the performance, contrarily, its performance

can get saturated due to vanishing gradient issues.

Thus, ResNet was introduced to overcome this issue

by adding identity shortcut connections and skipping

one or more convolutional layers as shown in Fig.3

(He et al., 2016a). Then the residual block has been

ameliorated generating a pre-activation variant of the

residual block where the gradient can ﬂow through

Figure 3: The residual block.

the shortcut connections to any other earlier convolu-

tional layer(He et al., 2016b).

Fig.4 illustrates the segmentation process using

the U-Net-based ResNet50 encoder.

Figure 4: Semantic segmentation result when using the U-

net architecture based ResNet50 encoder: Original image,

and segmentation map.

2.4 Majority Voting

After applying the three techniques separately, a fu-

sion method on the three segmentation maps is used.

The fusion of multiple result predictions has been

used by many researchers as an accurate strategy to

ameliorate the performance of many pattern recog-

nition systems, rather than relying on one single

model prediction Many techniques have been pro-

posed such as majority voting, bayesian decision

fusion(Yue et al., 2012), or Dempster-Shafter the-

ory(Kang et al., 2020) for fuzzy information fusion.

In this paper, we used the majority voting method

which is one of the basic and intuitive approaches.

It assigns a sample based on the most frequent class

assignment(Ballabio et al., 2019). For semantic seg-

mentation, each pixel in the output of the major-

ity voting process is assigned a background or fore-

ground class based on the three other correspond-

ing pixels of the three segmentation maps issued by

ICT4AWE 2025 - 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

372

thresholding, snakes, and U-Net-based ResNet.

3 EXPERIMENTS AND RESULTS

Experiments used to evaluate the performance of

DL models are not only an academic practice but

even when creating commercial systems, experiments

are necessary to create the most performant models.

Thus, to demonstrate the performance of our pro-

posed method, we have tested the methods separately

using different strategies and then we have reported

the result of fusing them. The experiments were con-

ducted using the ISIC2017 ( International Skin Imag-

ing Collaboration) dataset(Codella et al., 2018).

3.1 Technical Materials

Implementing these methods was done using Keras,

Sklearn, OpenCV, and TensorFlow frameworks. The

Keras and TensorFlow are used on a single machine

with an NVIDIA GEFORCE GTX-1050 Ti.

3.2 Dataset Representation

The experiments were conducted using a public

dataset which is collected from the ISIC archive, the

dataset was taken from a competition on skin lesion

segmentation that held in 2017(Codella et al., 2018).

The images were collected from a dermatoscopy tool,

and it contains about 2000 images of malignant and

benign lesions with their corresponding segmentation

maps. The dermoscopic lesion images are in JPEG

format, the size of almost all images is 1022×767

pixels. Their ground truth images are mask images

in PNG format. They are encoded as grayscale(8-

bit), where each pixel is 0 for pixels belonging to the

healthy skin(Background) and 255 for pixels belong-

ing to the suspicious lesion.

3.3 Evaluation Strategy

Our proposed method consists of fusing different

types of methods, those that are trainable and those

that use classical image processing methods. To

demonstrate the performance of our proposed method

the data set is split into two groups: a testing set con-

taining 600 images and 1400 images used for build-

ing the DL model. The selected test images are used

to test the DL model and to evaluate the performance

of the two other methods, as well. The normal pro-

cedure when evaluating a hypothesis is to validate it.

The validation helps to formulate the hypothesis to

see other proprieties and relations that can make it

better understood. Thus, this process will be itera-

tively repeated to ﬁnd a better version of the hypoth-

esis, also the same process is needed when creating

ML and DL models. While training the DL model,

a validation step should take place at each epoch. It

is necessary to validate its performance to report how

well the model is learning. Yet, a part of the train-

ing set will be held out for validation, and when the

experiments seem to be successful, the ﬁnal evalua-

tion will be done on the testing set. Therefore, split-

ting the training set into two sets one for training and

the other for validation may fail in accurately building

the best machine learning model due to the reduction

of the number of samples used for learning. Many

validation strategies are introduced for evaluating the

model(James et al., 2013), such as k-fold,leave-one-

out Cross-Validation, etc. The latter strategy consists

of validating one image using the model that has been

trained on the other remaining images. However, the

k-fold Cross-Validation is more interesting and rele-

vant since it minimizes the number of validation pro-

cesses by dividing the N data points into k subsets.

The training and the validation are then repeated k it-

erations, in each iteration one of the subsets is used

for validation and the remaining for training. The er-

ror is also averaged over all the k trials. Basically, k

can be either 5 or 10. Thus, we have used this strat-

egy for the validation of the DL model. We have also

evaluated its performance by splitting the dataset into

two sets: 1250 sample images for training and 150

for validation. This is the same conventional splitting

strategy used in(Codella et al., 2018).

3.4 Experimental Results

Fine-tuning the parameters depends on the nature of

the ML model. For deep learning, generally, the num-

ber of epochs will vary, simultaneously the training

and validation accuracy will be computed. This ex-

periment shows how the DL model is proceeding at

different epoch numbers. Regarding the experiments

done by Md Zahangir Alom et al.(Alom et al., 2018),

the training and validation accuracy becomes slightly

stable when using more than 50 epochs. Thus, this

number is adopted in this work.

We have trained the U-Net based Resnet encoder

with 1250 images, the 150 images will be used for

validation during training with a batch size of 12,

we have used the ADAM optimizer technique. Also,

the images were resized to 256 × 256 pixels and aug-

mented by applying many transformations such as ro-

tation, scale, color, and ﬂipping. Using these trans-

formations on the images will increase the size of

the dataset, automatically ameliorating the training

A Streamlined Lesion Segmentation Method Using Deep Learning and Image Processing for a Further Melanoma Diagnosis

373

Figure 5: Validation and training accuracy using U-net

based ResNet50.

Figure 6: Validation and training accuracy using

DeepLabv3+ based ResNet50.

performance. Fig5 shows both the training and val-

idation accuracy when using 50 epochs. The accu-

racy is stable for both validation and training. We

have also tested the DeepLabv3+ based Resnet50 en-

coder(Chen et al., 2018). The training and validation

accuracy are shown in Fig.6 which shows many oscil-

lations compared to Fig.5. It is caused by the use of

a small batch. It is observed that when using a small

batch for some models, it will appear a huge degra-

dation in the training process. Thus, the batch size

is an important hyperparameter that leads to either

sharp or ﬂat minimizers. We have used many strate-

gies for evaluating the performance of the proposed

method. For that, using the U-Net based ResNet50

encoder with a batch size of 12 is good enough for

accurately training the model and for minimizing the

amount of memory consumption. For a deeper study

of the U-net architecture, we have used more than

one validation strategy with 20 epochs. First, training

and validation were done using the conventional split-

ting strategy. Validation was performed using ran-

domly selected 150 images. Another strategy is the

k-fold cross-validation was performed. Using the k-

fold strategy can notably increase the diversity of data

used for both validation and training. The model will

Figure 7: Validation and training accuracy using U-net

based ResNet50 with different validation strategies.

be more generalized and unbiased. The training accu-

racy for both k-fold strategies is slightly higher than

using the conventional strategy (see Fig.7). However,

the validation accuracy is considerably greater when

using k-fold cross-validation for k=5 and k=10. Evi-

dently, models were stable in their predictions due to

the effectiveness of the training phase and the diver-

sity of validation samples.

After training the U-Net based Resnet50 model

using 1250 images during the 50 epochs, the network

weights are saved when the validation accuracy is at

its highest. These weights will be used then to make

predictions on the 600 testing images. Each one of

the three techniques will result in an output segmen-

tation mask. Table1 reports the quantitative results

of the experiments. The accuracy and IoU scores are

shown for each technique as well as for fusing them.

Fig.8 highlights some of the output examples of our

proposed method, it is obvious that the ﬁfth column

contains the most accurately segmented skin lesions

compared to the output of the DL model predictions.

Fig.8 shows that the DL model segments the le-

sion accurately when thresholding fails in segment-

ing it. However, using majority voting to fuse the

three techniques can reﬁne the segmentation result.

The irregularity is an important border-based feature

to recognize Melanoma, the proposed method can

perfectly draw attention to sharp borders contrarily

to the DL model. The ISIC2017 dataset has been

proposed in the 2017 skin lesion segmentation chal-

lenge. Yading Yuan has proposed a framework based

on a deep fully convolutional-deconvolutional neu-

ral network(CDNN) to automatically detect skin le-

sions(Yuan, 2017). This method yielded a Mean IoU

of 0.784 on the testing set, our proposed method pro-

vides a testing Mean IoU of 0.912 with a high accu-

racy of 0.983.

ICT4AWE 2025 - 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

374

Figure 8: Experimental results for skin lesion segmentation: (a) original lesion image, the segmentation map from: (b)

thresholding, (c) snakes, (d) U-net based Resnet50, (e) the proposed method, and (f) ground truth segmentation mask.

Table 1: Testing accuracy of the different methods and fusing them using majority voting.

Method Thresholding Snakes ResNet Majority Voting

Accuracy 87.61% 87.45% 96.64% 98.31%

Mean IoU 66.18% 66.66% 89.01% 91.29%

4 CONCLUSIONS

Statistics on melanoma incidence rates show that it

can be a non-life-threatening cancer if detected early.

Early diagnosis and treatment signiﬁcantly improve

the chances of successful recovery. The healthcare

community has been highly invested in creating au-

tomatic systems that can help doctors and patients

quickly and accurately inspect melanoma. These sys-

tems are built based on image processing and machine

learning techniques. One typical process is segment-

ing the skin lesion that will be inspected and then rec-

ognizing it as malignant or benign. In this work, we

have proposed a new method for lesion segmentation

that fuses three different methods: thresholding using

Otsu’s algorithm, morphological snake, and U-Net-

based Resnet50. Data fusion, speciﬁcally decision

fusion combines the decisions of multiple decisions

into a common one. The majority voting was used

to generate the ﬁnal segmentation map. Using other

decision fusion techniques can be adapted further to

ameliorate the performance of the proposed system.

The U-Net-based Resnet50 can be also explored us-

ing data fusion in the encoder or decoder units.

ACKNOWLEDGEMENTS

The authors would like to appreciably thank all the

members of LIS laboratory, for their direct help and

collaboration. This research was done during an in-

ternship at LIS laboratory, SeaTech, University of

Toulon.

REFERENCES

Alom, M. Z., Hasan, M., Yakopcic, C., Taha, T. M.,

and Asari, V. K. (2018). Recurrent residual con-

volutional neural network based on u-net (r2u-net)

for medical image segmentation. arXiv preprint

arXiv:1802.06955.

Armato, S. G., Li, F., Giger, M. L., MacMahon, H., Sone,

S., and Doi, K. (2002). Lung cancer: performance

of automated lung nodule detection applied to can-

cers missed in a ct screening program. Radiology,

225(3):685–692.

Ashwin, S., Ramesh, J., Kumar, S. A., and Gunavathi,

K. (2012). Efﬁcient and reliable lung nodule detec-

tion using a neural network based computer aided di-

agnosis system. In 2012 International Conference

on Emerging Trends in Electrical Engineering and

Energy Management (ICETEEEM), pages 135–142.

IEEE.

A Streamlined Lesion Segmentation Method Using Deep Learning and Image Processing for a Further Melanoma Diagnosis

375

Bae, K. T., Kim, J.-S., Na, Y.-H., Kim, K. G., and Kim,

J.-H. (2005). Pulmonary nodules: automated detec-

tion on ct images with morphologic matching algo-

rithm—preliminary results. Radiology, 236(1):286–

293.

Ballabio, D., Todeschini, R., and Consonni, V. (2019). Re-

cent advances in high-level fusion methods to classify

multiple analytical chemical data. Data Handling in

Science and Technology, 31:129–155.

Chan, T. and Vese, L. (1999). An active contour model

without edges. In International Conference on Scale-

Space Theories in Computer Vision, pages 141–151.

Springer.

Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., and

Adam, H. (2018). Encoder-decoder with atrous sepa-

rable convolution for semantic image segmentation. In

Proceedings of the European conference on computer

vision (ECCV), pages 801–818.

Codella, N. C., Gutman, D., Celebi, M. E., Helba, B.,

Marchetti, M. A., Dusza, S. W., Kalloo, A., Liopy-

ris, K., Mishra, N., Kittler, H., et al. (2018). Skin

lesion analysis toward melanoma detection: A chal-

lenge at the 2017 international symposium on biomed-

ical imaging (isbi), hosted by the international skin

imaging collaboration (isic). In 2018 IEEE 15th In-

ternational Symposium on Biomedical Imaging (ISBI

2018), pages 168–172. IEEE.

Daghrir, J., Tlig, L., Bouchouicha, M., and Sayadi, M.

(2020). Melanoma skin cancer detection using deep

learning and classical machine learning techniques: A

hybrid approach. In 2020 5th International Confer-

ence on Advanced Technologies for Signal and Image

Processing (ATSIP), pages 1–5. IEEE.

He, K., Zhang, X., Ren, S., and Sun, J. (2016a). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

He, K., Zhang, X., Ren, S., and Sun, J. (2016b). Iden-

tity mappings in deep residual networks. In Euro-

pean conference on computer vision, pages 630–645.

Springer.

Heinzerling, L. and Eigentler, T. K. (2021). Skin cancer

in childhood and adolescents: Treatment and implica-

tions for the long-term follow-up. In Late Treatment

Effects and Cancer Survivor Care in the Young, pages

349–355. Springer.

Jadhav, A. R., Ghontale, A. G., and Shrivastava, V. K.

(2019). Segmentation and border detection of

melanoma lesions using convolutional neural network

and svm. In Computational Intelligence: Theories,

Applications and Future Directions-Volume I, pages

97–108. Springer.

James, G., Witten, D., Hastie, T., and Tibshirani, R. (2013).

An introduction to statistical learning, volume 112.

Springer.

Jamil, U., Sajid, A., Hussain, M., Aldabbas, O., Alam, A.,

and Shaﬁq, M. U. (2019). Melanoma segmentation

using bio-medical image analysis for smarter mobile

healthcare. Journal of Ambient Intelligence and Hu-

manized Computing, 10(10):4099–4120.

Kang, B., Zhang, P., Gao, Z., Chhipi-Shrestha, G., Hewage,

K., and Sadiq, R. (2020). Environmental assessment

under uncertainty using dempster–shafer theory and

z-numbers. Journal of Ambient Intelligence and Hu-

manized Computing, 11(5):2041–2060.

Khazaei, Z., Ghorat, F., Jarrahi, A., Adineh, H., Sohrabi-

vafa, M., and Goodarzi, E. (2019). Global incidence

and mortality of skin cancer by histological subtype

and its relationship with the human development in-

dex (hdi); an ecology study in 2018. World Cancer

Res J, 6(2):e13.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., Van Der Laak, J. A.,

Van Ginneken, B., and S

anchez, C. I. (2017). A survey

on deep learning in medical image analysis. Medical

image analysis, 42:60–88.

Maglogiannis, I. and Doukas, C. N. (2009). Overview of ad-

vanced computer vision systems for skin lesions char-

acterization. IEEE transactions on information tech-

nology in biomedicine, 13(5):721–733.

Marquez-Neila, P., Baumela, L., and Alvarez, L. (2013). A

morphological approach to curvature-based evolution

of curves and surfaces. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 36(1):2–17.

Mbarki, W., Bouchouicha, M., Frizzi, S., Tshibasu, F.,

Farhat, L. B., and Sayadi, M. (2020). Semi-automatic

segmentation of intervertebral disc for diagnosing her-

niation using axial view mri. In 2020 5th International

Conference on Advanced Technologies for Signal and

Image Processing (ATSIP), pages 1–6. IEEE.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,

Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bern-

stein, M., et al. (2015). Imagenet large scale visual

recognition challenge. International journal of com-

puter vision, 115(3):211–252.

Schadendorf, D., van Akkooi, A. C., Berking, C.,

Griewank, K. G., Gutzmer, R., Hauschild, A., Stang,

A., Roesch, A., and Ugurel, S. (2018). Melanoma.

The Lancet, 392(10151):971–984.

Sharma, N. and Ray, A. K. (2006). Computer aided seg-

mentation of medical images based on hybridized ap-

proach of edge and region based techniques. In Proc.

Int. Conf. Math. Biol, pages 150–155.

Yuan, Y. (2017). Automatic skin lesion segmentation with

fully convolutional-deconvolutional networks. arXiv

preprint arXiv:1703.05165.

Yue, D., Guo, M., Chen, Y., and Huang, Y. (2012). A

bayesian decision fusion approach for microrna target

prediction. BMC genomics, 13(8):1–11.

ICT4AWE 2025 - 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

376