A Real-World Segmentation Model for Melanocytic and Nonmelanocytic

Dermoscopic Images

Eleonora Melissa

1 a

, Daria Riabitch

1 b

, Linda Lazzeri

, Federica La Rosa

1 c

, Chiara Benvenuti

Mario D’Acunto

3 d

, Giovanni Bagnoni

, Daniela Massi

4 e

and Marco Laurino

1 f

Institute of Clinical Physiology, National Research Council, Pisa, Italy

Uniti of Dermatologia, Specialist Surgery Area, Department of General Surgery, Livorno Hospital, Azienda Usl Toscana

Nord Ovest, Italy

Institute of Biophysics, National Research Council, Pisa, Italy

Department of Health Sciences, Section of Pathological Anatomy, University of Florence, Florence, Italy

Keywords:

Segmentation, Skin Lesions, Deep Learning, Border Features.

Abstract:

Segmentation is a critical step in computer-aided diagnosis (CAD) systems for skin lesion classiﬁcation. In

this study, we applied the Deeplabv3+ network to segment real dermoscopic images. The model was trained on

public datasets and tested both on public and on a disjoint set of images from the TELEMO project, covering

six clinically signiﬁcant skin lesion types: basal cell carcinoma, squamous cell carcinoma, melanoma, benign

nevi, actinic keratosis and seborrheic keratosis. Our model achieved a testing global accuracy of 97.88% on

public dataset and of 92.62% on TELEMO dataset, outperforming literature models. Although some misclassi-

ﬁcations occurred, largely due to class imbalance, the model demonstrated strong generalization to real-world

clinical images, a critical achievement for deep learning in medical imaging. To evaluate the clinical relevance

of our segmentation, we extracted ten key features related to lesion border and diameter. Notably, the ”Diame-

ters Mean” and ”Area to Perimeter Product” features revealed signiﬁcant differences between melanoma-nevi

and basal cell carcinoma-nevi classes, with strong effect sizes. These ﬁndings suggest that border features

are crucial for distinguishing between multiple skin lesion types, highlighting the model’s potential for aiding

dermatological diagnoses.

1 INTRODUCTION

Diagnosis of skin lesions has become increasingly im-

portant in clinical practice, especially when the early

diagnosis of skin tumors can increase patient survival

rates (Rojas et al., 2022). Melanoma poses signiﬁcant

challenges in clinical dermatology, as it is the deadli-

est form of skin cancer with a steadily increasing inci-

dence over recent decades. It results from the malig-

nant transformation of melanocytes, which can arise

from healthy skin or develop from congenital nevi.

Despite its aggressiveness, melanoma has favorable

survival rates when detected early. However, its iden-

tiﬁcation can be challenging, as it may be confused

https://orcid.org/0009-0005-8892-0470

https://orcid.org/0009-0007-7502-695X

https://orcid.org/0000-0003-1671-6057

https://orcid.org/0000-0003-4233-1943

https://orcid.org/0000-0002-5688-5923

https://orcid.org/0000-0003-4798-5196

with other nonmelanocytic skin cancers that present

similar dermoscopic features.

Nonmelanoma skin cancers (NMSCs) are more

prevalent, with a global incidence of 7.7 million cases

(Society, 2019). This category includes basal cell

carcinoma (BCC), squamous cell carcinoma (SCC),

and actinic keratosis (AK). BCC is a high survival

rate skin tumor (100% over 5 years (Society, 2019))

and is recognized as the most common malignant ep-

ithelial cancer. SCC is the second most widespread

nonmelanoma skin cancer behind BCC and it is of-

ten seen in elderly populations (et al., 2021; et al.,

2012). AK is a sun-induced precancerous lesion and

it may progress to SCC; its dermoscopic characteris-

tics present diagnostic challenges that require careful

evaluation (et al., 2021). Seborrheic keratosis (SK),

although it is a benign lesion, can mimic melanoma

and AK and complicates the differential diagnosis

with its varied color and unique structures (et al.,

2019; Hafner and Vogt, 2008).

316

Melissa, E., Riabitch, D., Lazzeri, L., La Rosa, F., Benvenuti, C., D’Acunto, M., Bagnoni, G., Massi, D. and Laurino, M.

A Real-World Segmentation Model for Melanocytic and Nonmelanocytic Dermoscopic Images.

DOI: 10.5220/0013129400003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 316-323

ISBN: 978-989-758-731-3; ISSN: 2184-4305

Recent advancements in artiﬁcial intelligence (AI)

have revolutionized the diagnostic landscape for skin

lesions. Automated tools can facilitate their identiﬁ-

cation, particularly skin tumors, making the diagnos-

tic process standardized and less operator dependent.

This advancement enhances the screening and early

detection of skin cancers, potentially improving pa-

tient outcomes and treatment success.

One of the fundamental steps of automatic skin

lesion classiﬁcation is segmentation, which poses an

important challenge in the ﬁeld of computer-aided di-

agnosis (CAD) systems due to the low contrast and

different quality of dermoscopic images. As a result,

many segmentation methods have been investigated

and developed over the years (Oliveira et al., 2016).

Thresholding-based methods, such as Otsu’s method

(Otsu, 1979) and Xu method (Xu et al., 1999), sepa-

rate lesion and background pixels choosing a thresh-

old based on the histogram intensity levels of the der-

moscopic image. In Canny’s edge detector (Canny,

1986) variations of the gradient magnitude are used to

identify the borders of the lesion. Chen-Vese method

(Chan and Vese, 2001) is based on active curves

which move toward the boundaries respecting an en-

ergy minimization principle. Region-based methods

(Oliveira et al., 2016) rely on identifying similar prop-

erties among neighbouring pixels to segment the input

image into different regions with a growing criterion.

Reddy et al (Reddy et al., 2022) employed such an ap-

proach in their work, leveraging pixel color and con-

trast similarity as the underlying principle for region

homogeneity.

Nevertheless, in recent years deep-learning ap-

proaches have shown better performances in terms of

skin lesion segmentation and border detection. The

earliest attempts to use deep networks for skin lesion

segmentation focused on the application of Convo-

lutional Neural Network (CNN) (Ronneberger et al.,

2015) and Fully Convolutional Network (FCN) (Bi

et al., 2018; Yuan et al., 2017), but then the U-Net

emerged as the leading deep network for biomedi-

cal image segmentations (Seeja and Suresh, 2019).

Many extensions of the U-Net, such as ResUNet, Re-

sUNet++, Fuzzy U-Net and attention U-Net (Ashraf

et al., 2022; Bindhu and Thanammal, 2023; Tong

et al., 2021), were eventually developed to obtain bet-

ter results, but also new nets were employed for this

task. Deeplabv3 (Wang et al., 2018) is a deep network

speciﬁcally designed for semantic segmentation and

its advanced version Deeplabv3+ has been used for

skin lesions segmentation combined with other archi-

tectures such as VGG19-Unet (Ali et al., 2019) and

MobileNetv2 (Zafar et al., 2023) with good results.

However, as far as we know, none of these studies has

tried to apply their models to real-world clinically ac-

quired images: training and testing are performed on

the same dataset usually split into two parts. Some-

times (Yuan et al., 2017), when a public dataset is

employed, the training and testing images are divided

by the database itself. Despite this, the quality of im-

ages collected from public databases differs signiﬁ-

cantly from that of images acquired during actual der-

matological inspections, which can affect the perfor-

mances of the methods.

In this context, we developed a Deeplabv3+ net-

work for the segmentation of real dermatoscopic im-

ages collected as part of the TELEMO project. We

trained the model using public datasets and tested it

both on a split portion of the public datasets and on

a completely disjoint dataset. Furthermore, to evalu-

ate the efﬁcacy of our model, we extracted ten of the

most relevant features related to border and diameter,

which are closely associated with segmentation qual-

ity. We then assessed whether these features could

effectively distinguish among six clinically relevant

types of skin lesions: basal cell carcinoma, squamous

cell carcinoma, melanoma, benign nevi, actinic ker-

atosis and seborrheic keratosis.

2 MATERIALS AND METHODS

2.1 Datasets

Our model was trained on different public datasets:

the ISIC 2017, ISIC 2019, ISIC 2020, PH2, 7-point

criteria evaluation database and a custom dataset by

Department of Dermatology of Hospital Italiano de

Buenos Aires (HIBA) (Lara et al., 2023), for a total

of 26597 images. A portion of the training dataset

(10% of the total images) was used to test the model

to enable comparison with literature models, which

utilized splitted datasets for training and testing. The

model was further tested on a separate test dataset col-

lected as part of the TELEMO project.

2.1.1 TELEMO Dataset

TELEMO (‘An Innovative TELEmedicine System for

the Early Screening of Melanoma in the General Pop-

ulation’) is a project whose primary objective is to de-

velop a teledermatology platform integrated with AI

tools for the analysis of skin lesions. A total of 479

dermoscopic images were obtained from patients with

an average age of 55 years (52 years for females and

57 years for males), using either the FotoFinder Medi-

acam 1000 (FotoFinder Systems GmbH, Germany) or

the Heine Delta 30 (HEINE Optotechnik, Germany).

A Real-World Segmentation Model for Melanocytic and Nonmelanocytic Dermoscopic Images

317

Figure 1: Comparison between public dataset images (ﬁrst

row) used as training set and TELEMO images (second

row) used as testing set.

The image quality produced by these clinical in-

struments, as previously highlighted, is very different

to that of public datasets (Figure 1), making the seg-

mentation task on real-world images more challeng-

ing. The skin lesion types of the TELEMO dataset

were distributed as follows: 53 melanoma (MEL),

232 benign nevi (NV), 20 squamous cell carcinoma

(SCC), 101 basal cell carcinoma (BCC), 15 actinic

keratosis (AK), 23 seborrheic keratosis (SK).

2.2 Image Segmentation

Table 1: This table summarizes the training options used for

the deep learning model.

Options Value

Backbone ResNet50

Optimizer sgdm

Max Epochs 30

Mini-batch size 16

Initial Learning Rate 0.001

L2 regularization 0.0005

Momentum 0.9

DeepLabv3+ is a network that uses a ResNet

backbone and integrates Atrous Spatial Pyramid

Pooling (ASPP) to capture multi-scale context, along

with a decoder module to reﬁne object boundaries

(Ali et al., 2019; Zafar et al., 2023). The net param-

eters used to train our Deeplabv3+ network in Mat-

lab®2023b are shown in Table 1. The ground-truth

segmentations for each image were already provided

online in the case of the ISIC 2017 and 7-point cri-

teria evaluation database. In the other cases, we ﬁrst

attempted to obtain reference segmentation using the

Topological Process Imaging method (Vandaele et al.,

2020), used in the work of Brutti et al (et al., 2023).

When this method proved unsuccessful, we manually

derived the segmentations with guidance from expert

dermatologists. The net required the resizing of all

the images to a standard dimension of 224x224 pixels.

Data augmentation techniques such as translation, ro-

tation, scaling, saturation, hue and brightness jittering

were applied to enhance the training set. The follow-

ing metrics were assessed to evaluate the performance

of our model:

1. Mean Accuracy: percentage of the correctly clas-

siﬁed pixels of all classes in all the images.

2. Global Accuracy: percentage of the correctly

classiﬁed pixels with respect to the total number

of pixels without the distinction of classes.

3. Mean IoU: also known as the Jaccard similarity

coefﬁcient, is evaluated as (1):

IoU index =

T P

T P + FP + FN

(1)

where TP is the number of true positive pixels, FP

is the number of false positive pixels and FN is the

number of false negative pixels.

4. Wheighted IoU: value of IoU for each class

weighted by the number of that class’ pixels.

5. Mean BF score: boundary F1 score, it indicates

how well the model predicted the boundaries of

the lesion for all the images.

2.3 Feature Extraction and Statistical

Analysis

The ABCD rule deﬁnes four primary feature classes

used to describe the appearance of skin lesions: asym-

metry, border, color, and diameter. These features

are regarded as the gold standard for distinguishing

melanoma from nonmelanoma skin lesions (Rao and

Ahn, 2012). In this study, we focused on the ﬁrst three

feature classes, which can be evaluated solely based

on the segmentation of dermoscopic images. Fur-

thermore, we explored whether these features could

also be applied to differentiate other types of skin le-

sions beyond melanoma. Hence we ﬁrst performed

the Levene’s test and the Shapiro-Wilk test to check if

the conditions of normality and homogeneity of vari-

ance were met for each feature. Since the results of

these tests showed a non-Gaussian distribution of the

data for every feature, we proceeded by performing

the Kruskall-Wallis test and the post-hoc Dunn test.

Additionally, for the signiﬁcant differences, we cal-

culated the epsilon squared index (ε

), an effect size

measure that indicates whether the differences also

have real clinical signiﬁcance.

2.3.1 Asymmetry

The asymmetry of the lesion is commonly measured

with the Normalized E-Factor (NEF), deﬁned by

Sancen-Plaza et al (et al., 2018) as (2):

NEF =

4 ×

√

(2)

BIOIMAGING 2025 - 12th International Conference on Bioimaging

318

where p is the perimeter of the lesion and A is its area.

2.3.2 Border

Khan et al. (Khan et al., 2020) identiﬁed three pri-

mary descriptors of lesion border quality with respect

to its irregularities:

1. Area to Perimeter Ratio (B1): ratio between the

area and the perimeter of the segmented lesion (3)

. (3)

2. Smooth index (B2): ratio between the area and the

square of the perimeter of the segmented lesion

(4)

4 ×π ×A

. (4)

3. Product of Area and Perimeter (B3): Product be-

tween the area and the perimeter of the segmented

lesion (5)

A ×p. (5)

Moreover, we also considered the fractal dimension,

an index of the lesion’s complexity and morphology,

which is calculated using the bounding-box algorithm

(Messadi et al., 2021; Ali et al., 2020).

2.3.3 Diameter

The real diameter of the lesion can’t be derived from

the dermoscopic images unless the magniﬁcation fac-

tor of the original image is known. Since this parame-

ter in not supplied for all the images, we derived other

descriptors of the lesion diameter which are (Khan

et al., 2020; Cavalcanti and Scharcanski, 2013; et al.,

2017):

1. Mean diameter (D1): mean between the two ap-

parent diameters (Da and Db) of the segmented

lesion (6)

+ D

with D

4 ×A

, D

D + d

(6)

where D is the major axis and d is the minor axis

of the segmented lesion.

2. Equivalent diameter: measure of diameter equiv-

alent to the real one: (7)

3 ×A

D × pi

. (7)

3. Axis Ratio: Ratio between the major (D) and mi-

nor (d) axis of the segmented lesion (8)

. (8)

4. Circularity: Index of the level of circularity of the

lesion (9)

4 × pi ×A

D × pi

(9)

3 RESULTS

3.1 Image Segmentation

Figure 2 (b) shows the confusion matrix for the public

images test dataset, while Figure 2 (a) corresponds

to the model performance on the TELEMO dataset.

The percentages in the confusion matrix refer to the

number of pixels which were correctly classiﬁed as

lesion pixels or background pixels.

In Table 2 the metrics obtained by our net are com-

pared with that of previous works.

3.2 Statistical Analysis

Figure 3 shows the box plots for all 10 features ex-

tracted for each class. The plot also includes brack-

ets with asterisks that connect the classes of fea-

tures showing signiﬁcant differences between multi-

ple pairs of skin lesions. The number of asterisks in-

dicates the level of signiﬁcance; the more asterisks,

the stronger the signiﬁcance.

Table 3 presents the epsilon squared values for

each signiﬁcant difference (indicated by the brackets

connecting classes in the plots of Figure 3) for each

feature.

4 DISCUSSION AND

CONCLUSIONS

Our application of the Deeplabv3+ network for image

segmentation yielded highly promising results, partic-

ularly in challenging cases where traditional methods

often struggle due to low contrast between lesions and

surrounding skin or the complex structure of lesion

boundaries.

This is evident in Figure 4: in image (a), the refer-

ence and Deeplabv3+ segmentations are nearly iden-

tical; in images (b) and (c), where lesion borders are

less clear and more complex, our segmentations show

only slight deviations from the reference. Greater in-

accuracies are observed in image (d), the most difﬁ-

cult case with very unclear boundaries. Nonetheless,

our segmentation remains quite similar to the refer-

ence, whereas other methods have completely failed

A Real-World Segmentation Model for Melanocytic and Nonmelanocytic Dermoscopic Images

319

(a) (b)

Figure 2: Testing confusion matrix of Deeplabv3+ net for TELEMO dataset (a) and public dataset (b).

Table 2: Performance of our model (internal testing on public dataset and external testing on TELEMO dataset) and previous

models.

Reference Method MA GA MIoU WIoU BF

(Ali et al., 2019) VGG19-Unet + Deeplabv3+ - 0.935 0.763 - -

(Zafar et al., 2023) MobileNetv2 + Deeplabv3+ 0.9147 0.9864 0.8814 0.9839 0.7836

(Vandaele et al., 2020) TPI** - 0.809 - - -

(Bi et al., 2018) FCN** - - 0.778 - -

(Tong et al., 2021) ASCU-Net** - 0.926 0.847 - -

(Yuan et al., 2017) FCN** 0.9555 - 0.847 - -

Our model Deeplabv3+ (internal testing) 0.9721 0.9788 0.9434 0.9589 0.9254

Our model Deeplabv3+ (external testing) 0.8895 0.9262 0.7901 0.8697 0.7553

*MA = Mean accuracy; GA = Global accuracy; MIoU = Mean IoU; WIoU = Weighted IoU; BF = Mean BF score.

** TPI = Topological Process Imaging; FCN = FUlly Convolutional Network; ASCU-Net = Attention Gate, Spatial and

Channel Attention U-Net.

Table 3: Features Kruskall-Wallis p-values and relatives ef-

fect size indexes.

Feature p-value ε

Asymmetry index 6.0940 x10

-5

0.0505

Area to perimeter ratio 3.1431x10

-7

0.0771

Area to perimeter product 1.8324x10

-11

0.1250

Smoothness index 1.1095 x10

-6

0.0708

Fractal dimension 0.0011 0.0351

Diameters mean 9.3245 x10

-11

0.1171

Equivalent diameter 8.9616 x10

-9

0.0948

Axis ratio 0.0626 0.0127

Circularity 0.0033 0.0294

in these cases. One notable issue is the misclassiﬁ-

cation of some background pixels as lesion pixels, as

our segmentations tend to be slightly larger than the

reference ones. This is also reﬂected in the confu-

sion matrices percentages (Figure 2) and can be at-

tributed to class imbalance, as the skin region domi-

nates most images with background pixels accounting

for an average of 76.20% and lesion pixels making

up 23.80%. This highlights the importance of using

more balanced datasets and paying close attention to

pixel-level class distributions in future reﬁnements.

Despite this limitation, our model achieved a

mean IoU of 94.34% and a mean accuracy of 97.21%

when tested on a split portion of the training dataset,

outperforming all the methods listed in Table 2. The

only metrics that are lower are global accuracy and

weighted IoU, compared to the second model in Table

2, which, however, is a more complex and computa-

tionally demanding network.

When tested on a disjoint real-world image

dataset, our model’s performance consistently de-

creased, although it still outperformed or remained

comparable to most of the models in Table 2. These

results underscore an important challenge in dermo-

scopic image segmentation: the need to generalize

models to real-world images. The quality of images

used for training, which are from public datasets, and

those captured in real clinical settings can differ sig-

niﬁcantly. This discrepancy can lead to a drastic drop

in model performances when applied to images that

deviate from the ones in public datasets. One possible

solution would be to use real dermoscopic images for

training; however, this could compromise the model’s

robustness due to the limited number of available im-

ages. Therefore, it is essential to use public datasets,

BIOIMAGING 2025 - 12th International Conference on Bioimaging

320

Figure 3: Features box plots. * stands for 0.01 < p −value < 0.05; ** stands for 0.001 < p −value < 0.01; *** stands for

p −value < 0.001.

A Real-World Segmentation Model for Melanocytic and Nonmelanocytic Dermoscopic Images

321

(a) (b)

Figure 4: Examples of some Deeplabv3+ segmentation

(blue line) overlapped to the reference segmentation (red

line) of TELEMO images.

which contain tens of thousands of images, during the

training phase, but the model must also be speciﬁcally

designed to adapt to varying image qualities, as has

been done in this work.

Given the accurate recognition of lesion borders

of our image segmentations, we would have expected

that at least a few border-related features would show

signiﬁcant differences between classes. As a matter

of fact, previous studies have shown that border and

color features are particularly useful for distinguish-

ing between skin lesions, especially melanoma and

common nevi (et al., 2017). This is conﬁrmed in our

ﬁndings, as only one feature, the axis ratio, has shown

a non-signiﬁcant p-value with also a small associated

effect size (Table3). From the boxplots of Figure 3,

the ”Diameters Mean” and ”Area to Perimeter Prod-

uct” features display two brackets with three asterisks,

indicating highly signiﬁcant differences between the

MEL-NV and BCC-NV classes. These two features

also have the highest epsilon squared index values,

suggesting a strong effect size and practical impor-

tance of the results (Table 3). Our statistical analysis

therefore suggests that features from the ABCD rule

are not only relevant in the identiﬁcation of melanoma

but also in distinguishing other skin tumors, such as

basal cell carcinoma.

For other features, even if the Kruskall-Wallis p-

values remain very low, the effect size is also moder-

ate or small. Additionally, all boxplots, except for

”Area to Perimeter Ratio”, show noticeable differ-

ences between the BCC and NV classes, suggesting

that border features are highly relevant in distinguish-

ing between these two classes. This warrants fur-

ther investigation with larger datasets, which could

also help determine whether the outlier values ob-

served, particularly in the ”Asymmetry Index”, ”Area

to Perimeter Product,” and ”Fractal Dimension”, are

due to the limited number of lesions analyzed in this

study or the high variability of these parameters in

dermoscopic images.

5 CONCLUSIONS

The Deeplabv3+ network developed in this work

for the segmentation of dermoscopic images demon-

strated high performance in handling challenging

cases where traditional methods often fail, particu-

larly when lesion boundaries are complex or indis-

tinct. This confers to our model a high generalization

ability to real-world clinical images, a crucial factor

in medical image analysis. Our accurate identiﬁca-

tion of lesion borders also enhances the distinction

between certain lesion classes, underscoring the di-

agnostic value of these features. Future work should

focus on reﬁning class balance and incorporating a

wider range of real clinical images to further vali-

date model performance. Furthermore, the potential

of border features for classifying various types of skin

lesions should be further explored by developing a

classiﬁcation model based on the extraction of these

features from the segmentation technique developed

in this work.

ACKNOWLEDGEMENTS

We thank Tuscany Region that funded the TELEMO

project, the ’Covid-19 Research Call Tuscany’ as

well as the THE-Tuscany Health Ecosystem grant

ECS

00000017.

REFERENCES

Ali, A. R., Li, J., Yang, G., and O’Shea, S. J. (2020). A

machine learning approach to automatic detection of

irregularity in skin lesion border using dermoscopic

images. PeerJ Computer Science, 6:1–35.

Ali, R., Hardie, R. C., Narayanan Narayanan, B., and

De Silva, S. (2019). Deep learning ensemble meth-

ods for skin lesion analysis towards melanoma detec-

tion. In 2019 IEEE National Aerospace and Electron-

ics Conference (NAECON), pages 311–316.

Ashraf, H., Waris, A., Ghafoor, M. F., Gilani, S. O., and Ni-

azi, I. K. (2022). Melanoma segmentation using deep

learning with test-time augmentations and conditional

random ﬁelds. Scientiﬁc Reports, 12.

Bi, L., Feng, D., and Kim, J. (2018). Improving automatic

skin lesion segmentation using adversarial learning

based data augmentation. CoRR, abs/1807.08392.

Bindhu, A. and Thanammal, K. K. (2023). Segmentation of

BIOIMAGING 2025 - 12th International Conference on Bioimaging

322

skin cancer using fuzzy u-network via deep learning.

Measurement: Sensors, 26.

Canny, J. (1986). A computational approach to edge detec-

tion.

Cavalcanti, P. G. and Scharcanski, J. (2013). Macro-

scopic pigmented skin lesion segmentation and its in-

ﬂuence on lesion classiﬁcation and diagnosis, vol-

ume 6, pages 15–39. Springer Netherlands.

Chan, T. F. and Vese, L. A. (2001). Active contours without

edges.

et al., A. S.-P. (2018). Quantitative evaluation of binary dig-

ital region asymmetry with application to skin lesion

detection philip payne. BMC Medical Informatics and

Decision Making, 18.

et al., C. R. (2012). Dermoscopy of squamous cell carci-

noma and keratoacanthoma. Archives of Dermatology,

148:1386–1392.

et al., F. B. (2023). Artiﬁcial intelligence algorithms for

benign vs. malignant dermoscopic skin lesion image

classiﬁcation. Bioengineering, 10.

et al., F. D. (2017). Segmentation and classiﬁcation of

melanoma and benign skin lesions. Optik, 140:749–

761.

et al., G. (2019). Atlante di dermoscopia. Piccin.

et al., V. R. (2021). Metabolomic analysis of actinic ker-

atosis and scc suggests a grade-independent model of

squamous cancerization. Cancers, 13.

Hafner, C. and Vogt, T. (2008). Seborrheic keratosis.

Khan, H., Yadav, A., Santiago, R., and Chaudhari,

S. (2020). Automated non-invasive diagnosis of

melanoma skin cancer using dermo-scopic images.

ITM Web of Conferences, 32:03029.

Lara, M. A. R., Kowalczuk, M. V. R., Eliceche, M. L.,

Ferraresso, M. G., Luna, D. R., Benitez, S. E., and

Mazzuoccolo, L. D. (2023). A dataset of skin lesion

images collected in argentina for the evaluation of ai

tools in this population. Scientiﬁc Data, 10.

Messadi, M., Cheriﬁ, H., and Bessaid, A. (2021). Segmen-

tation and abcd rule extraction for skin tumors classi-

ﬁcation.

Oliveira, R. B., Filho, M. E., Ma, Z., Papa, J. P., Pereira,

A. S., and Tavares, J. M. R. (2016). Computational

methods for the image segmentation of pigmented

skin lesions: A review.

Otsu, N. (1979). A threshold selection method from gray-

level histograms. IEEE Transactions on Systems,

Man, and Cybernetics, 9(1):62–66.

Rao, B. K. and Ahn, C. S. (2012). Dermatoscopy for

melanoma and pigmented lesions.

Reddy, D. A., Roy, S., Kumar, S., and Tripathi, R. (2022). A

scheme for effective skin disease detection using opti-

mized region growing segmentation and autoencoder

based classiﬁcation. In Procedia Computer Science,

volume 218, pages 274–282. Elsevier B.V.

Rojas, K. D., Perez, M. E., Marchetti, M. A., Nichols, A. J.,

Penedo, F. J., and Jaimes, N. (2022). Skin cancer:

Primary, secondary, and tertiary prevention. part ii.

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

Convolutional networks for biomedical image seg-

mentation. CoRR, abs/1505.04597.

Seeja, R. D. and Suresh, A. (2019). Deep learning

based skin lesion segmentation and classiﬁcation of

melanoma using support vector machine (svm). Asian

Paciﬁc Journal of Cancer Prevention, 20:1555–1561.

Society, C. C. (2019). Canadian cancer society.survival

statistics for non-melanoma skin cancer. [Accessed

12 August 2019].

Tong, X., Wei, J., Sun, B., Su, S., Zuo, Z., and Wu, P.

(2021). Ascu-net: Attention gate, spatial and channel

attention u-net for skin lesion segmentation. Diagnos-

tics, 11.

Vandaele, R., Nervo, G. A., and Gevaert, O. (2020). Topo-

logical image modiﬁcation for object detection and

topological image processing of skin lesions. Scien-

tiﬁc Reports, 10.

Wang, Y., School, T. H., Sun, S., and School, J. Y. T. H.

(2018). Skin lesion segmentation using atrous convo-

lution via deeplab v3.

Xu, L., Jackowski, M., Goshtasby, A., Roseman, D., Bines,

S., Yu, C., Dhawan, A., and Huntley, A. (1999). Seg-

mentation of skin cancer images. Image and Vision

Computing, 17(1):65–74.

Yuan, Y., Chao, M., and Lo, Y. C. (2017). Automatic

skin lesion segmentation using deep fully convolu-

tional networks with jaccard distance. IEEE Trans-

actions on Medical Imaging, 36:1876–1886.

Zafar, M., Amin, J., Sharif, M., Anjum, M. A., Mal-

lah, G. A., and Kadry, S. (2023). Deeplabv3+-based

segmentation and best features selection using slime

mould algorithm for multi-class skin lesion classiﬁca-

tion. Mathematics, 11.

A Real-World Segmentation Model for Melanocytic and Nonmelanocytic Dermoscopic Images

323