Image Segmentation using Gradient-based Histogram Thresholding

for Skin Lesion Delineation

Pedro M. M. Pereira

1,3

, Luis M. N. Tavora

, Rui Fonseca-Pinto

, Rui Pedro Paiva

3,4

Pedro A. A. Assuncao

1,2

and Sergio M. M. de Faria

1,2

Instituto de Telecomunicac¸˜oes, Portugal

ESTG, Polytechnic Institute of Leiria, Portugal

DEI - FCTUC, University of Coimbra, Portugal

CISUC, Department of Informatics Engineering, University of Coimbra, Portugal

Keywords:

Segmentation, Skin Lesion Detection, Medical Imaging, Dermoscopy.

Abstract:

Image segmentation is a key stage in medical image processing algorithms and machine learning classiﬁers

where identiﬁcation of discriminative features are of utmost importance. In the case of skin lesions, most of the

existing image segmentation approaches aim at minimising some error metric between computed and ground-

truth regions of interest (ROI) deﬁned by medical experts, where ROI delineation is not always considered.

This paper proposes an image segmentation method for skin lesion delineation, which expands traditional

histogram and clustering-based approaches to achieve the best trade-off between both. The proposed method

is capable of providing accurate details of the skin lesion borders, without deviating from the coarser borders

of the available ground-truth.

1 INTRODUCTION

In general, medical image processing systems include

a segmentation stage to identify regions of interest

(ROI) for further processing, which may include tex-

ture and colour analysis, feature extraction, etc. In the

case of pigmented skin lesions, due to the rather lim-

ited human capability to discriminate slight variations

in contrast and blur, precise identiﬁcation of relevant

ROI boundaries poses a problem to dermatologists

(Claridge and Orun, 2002). Morphological aspects

of skin lesions alongside with the large number of

environment-variables (e.g. location in the body, skin

properties, lighting conditions and angle of view) fur-

ther increase the challenge of accurate segmentation

of the most useful ROI (Celebi et al., 2009a). This re-

sults in signiﬁcant inter and intra-observer variability

and coarse ROI segmentation. Thus, to reduce the de-

pendence of human factors, different types of compu-

tational methods have been used for image segmenta-

tion, spanning over a quite considerable range of cat-

egories (Pathan et al., 2018).

Recent advances in machine learning approaches

are rapidly changing the landscape of medical image

processing algorithms for detection, recognition and

classiﬁcation, where data sets with accurate ground-

truth image segmentation are increasingly necessary

both for training and validation of such new com-

putational models (Ker et al., 2018; Oliveira et al.,

2018). In the case of skin lesion segmentation, the dif-

ﬁculty of achieving accurate delineation of ROI bor-

ders manually, has driven research efforts to increase

the availability of ground-truth ROI through com-

putational methods (Cheng et al., 2015; K´echichian

et al., 2014). Particularly relevant in the applica-

tion scope of this work is the mobile system designed

for early detection of melanoma recently proposed in

(Do et al., 2018), which combines fast segmentation,

feature extraction and classiﬁcation in resource con-

strained devices.

This paper proposes an accurate segmentation

method for pigmented skin lesions, envisaging delin-

eation of melanoma as the main application. In gen-

eral this kind of medical images produce bi-modal

histograms, and although this characteristic has been

used as the basis of different segmentation methods, it

results in either coarse borders or simply fails to pro-

vide signiﬁcant ROI in images with low colour con-

trast and smooth texture transitions. Therefore, this

work addresses the problem of accurate identiﬁcation

of the relevant ROI in such images, which includes the

ability to deﬁne the external border of the lesion with

Pereira, P., Tavora, L., Fonseca-Pinto, R., Paiva, R., Assuncao, P. and M. de Faria, S.

Image Segmentation using Gradient-based Histogram Thresholding for Skin Lesion Delineation.

DOI: 10.5220/0007354100840091

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 84-91

ISBN: 978-989-758-353-7

high level of precision. Taking into account the clin-

ical relevance of this aspect, a gradient-based metric

is devised to drive the proposed delineation method

across a reﬁnement histogram-based segmentation al-

gorithm. Two different types of medical images are

targeted, which increases the challenge of achieving

efﬁcient segmentation with accurate border details in

both cases, as pointed out in (Zhou et al., 2008). For

this purpose, the Dermoﬁt dataset of macroscopic im-

ages (Ballerini et al., 2013) and the PH

dataset of

dermoscopic images (Mendonc¸a et al., 2013) are both

used in this work.

This work is organised as follows: Section 2

presents the background that is relevant for the pro-

posed method. Section 3 describes the Gradient-

based metric and Section 4 introduces the proposed

method. In Section 5 the obtained experimental re-

sults are presented, alongside with the associated dis-

cussion. Finally, in Section 6 some conclusions and

future work perspectives are presented.

2 SKIN LESION SEGMENTATION

- BACKGROUND

As mentioned above, images of skin lesions exhibit

two distinctive regions which are associated, one to

the lesion itself, and another to the surrounding skin.

This leads to bi-modal histograms, as shown in Fig. 1,

for images of skin lesions with quite different charac-

teristics. This results not only from the lesion mor-

phology, but also from the use of different image ac-

quisition technologies and lighting conditions. For in-

stance, while accurate segmentation of the skin lesion

shown in Fig. 1a is quite easy to be obtained directly

from its well-deﬁned histogram, in the case of Fig. 1c

it poses a problem due to the blurry borders and low

colour contrast. This can be conﬁrmed by the corre-

sponding histogram shown on the right side of Fig. 1,

where the pixels belonging to the relevant ROI are

quite difﬁcult to identify.

2.1 Histogram Thresholding

Histogram thresholding techniques have long been

used for segmentation of these type of images, where

the region of the lesion can be distinguished by its dif-

ferent tonality (Korotkov and Garcia, 2012). The un-

derlying idea of these methods is to perform a binary

partition of the image based on the luminance level

of each pixel, meaning in this case (e.g. Fig. 1) to

successfully separate the region of the lesion (darker

region ↔ left Y-peak: Y

Pmin

) from the surrounding

skin (right Y-peak: Y

Pmax

). In a simple formulation,

the segmentation challenge to be considered here is to

ﬁnd an optimum criterion to deﬁne a threshold value

) that leads to an accurate ROI extraction, i.e., the

region of the image that contains the lesion. Different

threshold techniques exist for decades (Sahoo et al.,

1988), and this operation can be done either directly

on the Y-histogram or after a transformation T(Y), as

proposed in (Rajab et al., 2004). Nevertheless, the

performance of the method might strongly depend on

the distribution of luminance values, as inferred from

Fig. 1c and even Fig. 1b.

2.2 Clustering

Clustering algorithms have also been used for

skin lesion segmentation based on different ap-

proaches (Melli et al., 2006; Tasoulis et al., 2010;

Iyatomi et al., 2008; Ganster et al., 2001). The al-

gorithms can be fed with image information in differ-

ent formats such as RGB, YUV or YCbCr, but there

are also systems using only the luminance Y channel

since the inherent fusion process of the RGB chan-

nels allows inclusion of the relevant colour informa-

tion (Maglogiannis and Doukas, 2009). A quite ef-

ﬁcient clustering approach is based on the so called

K-Means, or Lloyd’s algorithm (Lloyd, 1982), which

is an iterative data-partitioning algorithm that assigns,

for a predeﬁned number of clusters, every input ob-

servation to only one of the clusters. In skin lesion

images, clustering algorithms may take advantage of

the bi-modal characteristics of the histogram to use

the corresponding peaks as the initial centroids.

2.3 Other Approaches

Several other segmentation approaches are worth-

while to be mentioned. For instance, the concept of

Fuzzy Differential Evolution Entropy (FT) was used

in (Sarkar et al., 2014), through an algorithm that cre-

ates fuzzy partitions from the image histogram. Then

the entropy is optimised to obtain the thresholds us-

ing a differential evolution meta-heuristic. Another

algorithm based on a similar approach is the Fuzzy

Clustering LevelSet (FL) (Li et al., 2011), which

uses a hybrid model that alternates between global

and local region competitions using spatial informa-

tion in the fuzzy clustering technique. For Quantiza-

tion approaches one can ﬁnd early methods based on

PCT (Principal Components Transform) Median Cut

(PM) (Umbaugh et al., 1993), while Active Contours

were proposed in (Chan et al., 2000) and (Lankton

and Tannenbaum, 2008), referred as Chan-Vese (CV)

and Lankton Mean Separation (LM). In the former a

Mumford-Shah function is minimised over the length

Image Segmentation using Gradient-based Histogram Thresholding for Skin Lesion Delineation

100

150

200

250

0.5

1.5

·10

(a) A121a from (Ballerini et al., 2013)

100

150

200

250

0.2

0.4

0.6

0.8

·10

(b) IMD390 from (Mendonc¸a et al., 2013)

100

150

200

250

·10

Figure 1: Skin lesion images (left) and the corresponding

luminance (Y) histograms (right), were red dots represent

peaks that correspond to the lesion and skin, respectively.

of the contour, while the latter uses local image statis-

tics and evolving contours based on local information.

Additionally, traditional methods like Otsu Threshold

(OT) (Otsu, 1979) and K-Means Colour (KC) (Melli

et al., 2006) also serve as baseline literature.

3 GRADIENT-BASED METRIC

The ﬁgure of merit that is herein proposed to assess

the accuracy of a given outside border line is based

on the rationale that the segmentation contour sep-

arates regions with substantially different tonalities.

Accordingly, the magnitude of the image gradient on

contour pixels is expected to yield higher values than

in other regions (i.e., either inside the lesion or in the

remaining skin). Following this argument, segmenta-

tion masks whose border lines exhibit higher gradient

magnitudes should more faithfully separate the two

regions of the image.

For a given image I, the gradient magnitude in

each pixel can be determined by Eq.(1),





∂I

∂x





∂I

∂y



1/2

, (1)

ˆn

Figure 2: Segmentation border line (red) and gradient: the

higher the projection of the image gradient

G onto ˆn

, the

more accurate is the border line (Image P348a from (Bal-

lerini et al., 2013)).

where j denotes the colour (or luminance) channel

under consideration (e.g., j = R,G,B,Y). For any

point of a border line deﬁned by segmentation, the

projection of the gradient vector

G onto the orthogo-

nal direction of the line (ˆn

), deﬁnes an accuracy met-

ric for the border line. Such projection is given by (2).

⊥

l, j

· ˆn

(2)

This concept is depicted in Fig. 2, where one can

observe that the orthogonal direction of the segmen-

tation border line (red) is not aligned with the gradi-

ent at the same point. The higher the projection G

⊥

l, j

computed by (2) the better (i.e. more accurate) is the

lesion contour segment.

Therefore, following the above discussion, the av-

erage value of G

⊥

l, j

over all points of a contour line

l is used as the gradient-based metric to evaluate how

accurately a given segmentation contour represents

the outside border on a skin lesion.

4 PROPOSED METHOD

The image segmentation method proposed for skin le-

sion delineation follows the processing chain depicted

in Fig. 3. The underlying idea is to ﬁnd an optimal

ROI delineation based on a trade-off between an ROI

with the highest gradient magnitude in the orthogo-

nal direction of its border line, and another ROI with

larger area but lower gradient. While the former iden-

tiﬁes the sharpest boundary of the skin lesion, the lat-

ter contains more boundary information which is also

useful for medical analysis and monitoring of tempo-

ral evolution. As described next, gradient-based his-

togram thresholding and clustering are used to gener-

ate the two ROIs for ﬁnal optimisation and delineation

of skin lesions.

BIOIMAGING 2019 - 6th International Conference on Bioimaging

4.1 Histogram Thresholding

The bi-modal characteristic of skin lesion image his-

tograms is used to determine the two most impor-

tant peaks, which in turn deﬁne the range limits

Pmin

Pmax

for all possible thresholds, i.e., the best

ROI must be found by cutting the histogram at the

optimum threshold Y

∗

∈ {Y

Pmin

Pmax

} to be found

between the two peaks (e.g. red markers on the his-

tograms of Fig. 1). In the method shown in Fig. 3,

histogram thresholding is performed for all values be-

tween Y

Pmin

and Y

Pmax

, generating an equal number

of images and the corresponding segmentation masks.

After a ﬁltering process to remove small isolated re-

gions and outliers, a clean ROI is determined for each

image and the average gradient G

⊥

is computed, as

deﬁned by (2). Then the ROI whose border yields

the maximum average gradient is selected along with

its histogram threshold. In summary, this process en-

sures that the border line of such ROI is the one with

the highest tonality variations across it. In the remain-

ing sections, the ROI obtained by maximising the gra-

dient border through histogram thresholding is identi-

ﬁed by HT.

4.2 Clustering

In the proposed method, clustering is used to identify

a coarse ROI for the skin lesion where the boundaries

include, in general, the smooth transition regions be-

tween the lesion and the surrounding healthy skin. In

this work the variant K-Means++ was selected, due

to its faster clustering convergence and also good dis-

criminative performance due different heuristics used

for ﬁnding centroids (Arthur and Vassilvitskii, 2007).

The iterative clustering process is carried out with a

maximum of 200 iterations seeking for two clusters

with a global minimum of the Euclidean distance to

cluster-centre. For the sake of reproducibility, the

initial centroids are deﬁned as the histogram peaks.

In the remaining sections, the ROI obtained through

clustering is identiﬁed by KM.

4.3 Filtering

Due to noise, inherent illumination variations and

other factors, both the histogram thresholding and

clustering methods described above produce ROIs

with binary masks that include not only a large blob

(the skin lesion region) but also other small isolated

regions spread across the whole image. The ﬁlter-

ing process devised to remove such unwanted regions

assumes that the lesion region limits are fully lo-

cated within the image, so the ﬁrst operation is to

remove all isolated regions with any boundary coin-

cident with the image borders. This is done by us-

ing a ﬂood-ﬁll algorithm based on morphological re-

construction (Soille, 2013). This ﬁrst cleansing op-

eration is especially relevant when processing images

from the PH

dataset, as they exhibit a black circu-

lar frame artiﬁcially introduced in the dermatoscope

digitisation process. The relevant ROI containing the

lesion is then determined by extracting the blob with

the largest area in the binary mask, using a labelling

procedure (Haralick and Shapiro, 1992, p. 40-48).

4.4 Optimised Segmentation

The optimisation step aims at improving delineation

of skin lesions by selectively expanding the ROI

border that was previously found through histogram

thresholding with gradient maximisation, in order to

further include relevant areas of transition regions.

This is necessary because gradient maximisation of-

ten leads to stringent contour lines which only in-

clude the inner parts of the lesion and leave out rel-

evant transition regions. Taking into account that

clustering-based segmentation usually results in the

inclusion of a larger region around the inner part

of the lesion, the proposed optimisation procedure

achieves the best trade-off between gradient maximi-

sation and increased ROI area to include transition re-

gions. This is done by ﬁnding an optimum threshold

∗

) that maximises, simultaneously, both the gradi-

ent of the border line G

⊥

and the ROI area.

For each ROI obtained through histogram thresh-

olding (HT) and clustering (KM), as described above,

let us deﬁne the following reference values:

• G

⊥,KM

and G

⊥,HT

: the average gradient of the

border (in (2));

• A

and A

: the area of the ROI, i.e. skin lesion.

The corresponding values associated to an arbitrary

histogram thresholdY

are deﬁned as G

⊥,Y

and A

respectively. The ratios R

) and R

) (expres-

sion (3)) deﬁne the relative gradient and relative area

of any ROI obtained with threshold Y

, using the HT

and KM ROIs as references.

) =

⊥,Y

⊥,KM

) =

min(A

)

(3)

The optimisation procedure consists in ﬁnding the

optimum threshold Y

∗

that maximises both R

)

and R

). This is accomplished by maximising

their product, provided that the selected maximum

does not lead to gradient values below that of the

KM ROI (G

⊥,KM

) and the new ROI area falls between

Image Segmentation using Gradient-based Histogram Thresholding for Skin Lesion Delineation

Image

Histogram

Peak

Detection Thresholding

Segmentation

Filter

Cut

max(G

⊥

)

⊥,HT

⊥,KM

ROI

Clustering

Optimization

Final ROI

Figure 3: Proposed scheme.

those of HT and KM ROIs. This is equivalent to solve

the following constrained maximisation problem:

∗

= arg max

∈Y

) (4)

subject to:

≥ 1∧ R

≥ 1∧

max(A

)

≤ 1 (5)

In summary, the optimal histogram threshold Y

∗

to be used for delineation of skin lesions is found

through a trade-off between border gradient and

amount of transition area included in the ROI.

5 EXPERIMENTAL RESULTS

AND DISCUSSION

The performance of the segmentation algorithm de-

scribed in Section 4 was evaluated using sets of im-

ages from different databases. A total of 195 images

from the Dermoﬁt dataset (Ballerini et al., 2013) and

32 images from the PH

dataset (Mendonc¸a et al.,

2013) were used. This selection followed two main

criteria: i) images without hair strands crossing the

lesion, i.e., as hairless as possible and ii) lesion limits

within the image, i.e., the whole lesion boundary fully

located inside the image.

In the ﬁrst stage the input grayscale image passes

through two segmentation processes, namely His-

togram Thresholding and Clustering Segmentation.

As pointed out in section 4, the output of both algo-

rithms may exhibit some image artefacts, which can

be observed in Fig. 4b and Fig. 4e. Then the efﬁ-

ciency of the ﬁltering stage, that is used after both

segmentation algorithms (described in Section 4.3),

in removing the small isolated regions, is shown in

Fig. 4c and Fig. 4f. In these images it is possible to

observe that the ﬁltering process is quite effective in

providing an accurate lesion/skin segmentation mask

without harming the border details.

After the segmentation and ﬁltering stages, accu-

rate ROI delineation is performed, following the opti-

misation procedure described in the previous section.

(a) Image B964 (b) Histogram TH (c) Filtered

(d) image B241 (e) Clustering (f) Filtered

Figure 4: Image segmentation using Histogram thresh-

olding (HT) and K-Means (KM) clustering: (a,d) original

grayscale images from (Ballerini et al., 2013), (b,e) seg-

mentation output, (c,f) ﬁnal binary mask after the ﬁltering

operation.

For visual evaluation and discussion, a set of rep-

resentative types of skin lesions have been selected

from the datasets to represent the segmentation re-

sults, which is depicted in Fig. 5.

In Fig. 5 the lesion segmentation using KM is rep-

resented by a red line, the HT by a green line and the

Proposed method by a blue line. The white line rep-

resents the ground-truth (GT) provided by the dataset.

From the representative results presented in

Fig. 5a to Fig. 5e, it can be observed that the algo-

rithms are in general quite effective in the segmenta-

tion of images and delineation of the relevant ROI.

The Histogram Thresholding method (HT) is able

to achieve accurate delineation when there is a sharp

tonality difference between the skin lesion and the

surrounding skin. However, as mentioned before, in

images with smoother lesion-to-skin transitions, the

highest value of G

⊥

, may result in a segmented re-

gion that is smaller than expected. This effect can be

seen in images from Fig. 5f to Fig. 5l. This kind of

output is not the most useful from the clinical point of

view, as it may exclude a relevant part of the lesion.

In the case of Clustering ROI segmentation (KM),

BIOIMAGING 2019 - 6th International Conference on Bioimaging

(a) D427b (b) D155b (c) B447a

(d) D384 (e) A121a (f) B311c

(g) A92 (h) B17a (i) B964

(j) B1075 (k) IMD103 (l) IMD175

Figure 5: Skin lesion segmentation using KM (red), HT

(green) and Proposed (blue). The white line corresponds

to the dataset provided ground-truth (GT). Images (a)

to (j) are from (Ballerini et al., 2013), and (k) and (l)

from (Mendonc¸a et al., 2013).

in general, the segmented region may include smooth

transition regions between the lesion and the sur-

rounding healthy skin. This commonly results in

a larger region than that obtained by the Histogram

Threshold method, as can be seen in images Fig. 5h,

Fig. 5i, Fig. 5j and Fig. 5l. In such cases, this might

not represent the best option as well.

In order to overcome the HT underestimation of

the ROI and the possible KM overestimation, the pro-

posed combined method relies on a trade-off between

the ROI and the border gradient. As can be observed

in all images of Fig. 5, the blue line always represents

a more precise delineation of the ROI.

Some authors compare the segmentation results

with the ground-truth (GT) segmentation masks pro-

vided in the databases. Nevertheless, as can be visu-

ally observed in Fig. 5a, Fig. 5b and Fig. 5f, the GT

borders are not as accurate and spatially detailed as

those obtained with the used algorithms. It can also

be observed that the GT lines often miss areas with

high texture variations.

In numerical terms, other performance indicators

are usually considered as benchmarks (Hance et al.,

1996; Celebi et al., 2009b; Garnavi et al., 2011):

• Border Error (BE), in (6), a measure of the

fraction of non-overlapping (exclusive-OR, ⊕)

segmentation regions (Area) between the pro-

posed segmentation (PS) method and the dataset

ground-truth segmentation (GT);

BE(PS, GT) =

Area(PS⊕ GT)

Area(GT)

(6)

• True Detection Rate (TDR), in (7), an indicator of

the ratio of number of pixels (np) that are correctly

classiﬁed as lesion;

TDR(PS, GT) =

np(PS ∩ GT)

np(GT)

(7)

• False Positive Rate (FPR), in (8), which accounts

for the number of pixels that are incorrectly clas-

siﬁed as lesion.

FPR(PS,GT) =

np(PS ∩

GT)

np(GT)

(8)

The results obtained for these indicators are pre-

sented in Table 1, alongside with those from the meth-

ods presented in Section 2.3. It can been seen that the

performance of the proposed algorithm (Prop) is gen-

erally inline with others published in the literature,

though not always consistent for all metrics. How-

ever, it should be kept in mind that these indicators

use the GT as reference, which does not provide seg-

mentation masks with as much spatial details as those

herein obtained. Such difference can be clearly ob-

served in Fig. 5

The gradient metric deﬁned in Section 3 was

also used to assess the performance of the proposed

method. The quotient of gradient between delin-

eations for both datasets was determined for such pur-

pose and the results are presented in Table 2. Observ-

ing its ﬁrst three lines, it can be seen that HT has on

average the highest G

⊥

values, as the method was op-

timised for such purpose, though in some cases this

also corresponds to inaccurate segmentation. The sec-

ond group of three lines make it clear that the Prop

method outperforms KM while only slightly compro-

mising the G

⊥

value in comparison with the max-

imum of HT. The remaining data on the table also

shows that the Prop method produces segmentations

Image Segmentation using Gradient-based Histogram Thresholding for Skin Lesion Delineation

Table 1: Ground Truth (GT)-based indicators (%).

Method

Dermoﬁt PH2

BE TDR FPR BE TDR FPR

OT 32.569 76.386 7.483 20.907 83.999 7.061

KC 53.475 78.439 8.581 18.174 86.041 6.313

FT 1.017 37.819 25.731 50.499 50.734 15.319

FL 2.439 49.500 50.084 20.352 82.846 6.872

PM 1.504 97.456 23.096 1.107 93.442 24.524

CV 2.383 69.381 35.618 48.521 68.161 14.119

LM 2.164 0.0003 52.467 1.604 94.543 36.550

HT 36.253 64.920 9.409 34.763 65.586 11.165

KM 21.592 79.336 5.554 17.732 85.113 6.085

Prop 23.034 78.283 6.008 20.336 81.676 6.950

with G

⊥

values higher than any of the other algo-

rithms previously introduced. This means that skin

lesion delineation obtained by the proposed method

is more accurate than the others because the border

line is found where the gradient is higher, i.e., a bet-

ter discrimination between lesion and normal skin is

obtained.

Table 2: Average Border Gradient indicators.

Indicators Dermoﬁt PH2

⊥,HT

/ G

⊥,KM

1.173 1.271

⊥,Prop

/ G

⊥,HT

0.933 0.917

⊥,Prop

/ G

⊥,KM

1.082 1.086

⊥,HT

/ G

⊥,GT

3.908 4.742

⊥,KM

/ G

⊥,GT

3.364 3.997

⊥,Prop

/ G

⊥,GT

3.688 4.301

⊥,Prop

/ G

⊥,OT

1,831 1,916

⊥,Prop

/ G

⊥,KC

1,831 1,921

⊥,Prop

/ G

⊥,FT

2,268 1,787

⊥,Prop

/ G

⊥,FL

6,429 1,829

⊥,Prop

/ G

⊥,PM

2,894 2,701

⊥,Prop

/ G

⊥,CV

1,175 1,086

⊥,Prop

/ G

⊥,LM

3,376 3,191

6 CONCLUSIONS

This paper addressed the segmentation of skin lesion

images using both histogram thresholding and clus-

tering algorithms to overcome the limitations of each

one on its own. A gradient-based method was devised

for optimised thresholding and ROI border quality pa-

rameter. The segmentation masks obtained for the ﬁ-

nal ROIs indicate that this method is quite accurate in

delineation of the relevant lesion regions containing

for a wide range of images. The experimental vali-

dation, using two publicly available images datasets,

show that the proposed approach is effective in delin-

eating skin lesions with detailed geometry in regions

with quite diverse tonality variations. The accurate

delineation of skin lesions is a relevant achievement

to provide discriminative features for machine learn-

ing algorithms and also to investigate patterns of tem-

poral evolution of the borders.

ACKNOWLEDGEMENTS

This work was supported by the Fundac¸˜ao para a

Ciˆencia e Tecnologia, Portugal, under PhD Grant

SFRH/BD/128669/2017 and project PlenoISLA

PTDC/EEI-TEL/28325/2017, in the scope of R&D

Unit 50008, through national funds and where

applicable co-funded by FEDER – PT2020.

REFERENCES

Arthur, D. and Vassilvitskii, S. (2007). k-means++: The

advantages of careful seeding. In Proceedings of the

eighteenth annual ACM-SIAM symposium on Discrete

algorithms, pages 1027–1035. Society for Industrial

and Applied Mathematics.

Ballerini, L., Fisher, R. B., Aldridge, B., and Rees, J.

(2013). A color and texture based hierarchical k-nn

approach to the classiﬁcation of non-melanoma

skin lesions. In Color Medical Image Analysis,

pages 63–86. Springer. https://licensing.edinburgh-

innovations.ed.ac.uk/i/software/dermoﬁt-image-

library.html.

Celebi, M. E., Iyatomi, H., Schaefer, G., and Stoecker, W. V.

(2009a). Lesion border detection in dermoscopy im-

ages. Computerized medical imaging and graphics,

33(2):148–153.

Celebi, M. E., Schaefer, G., Iyatomi, H., Stoecker, W. V.,

Malters, J. M., and Grichnik, J. M. (2009b). An im-

proved objective evaluation measure for border detec-

tion in dermoscopy images. Skin Research and Tech-

nology, 15(4):444–450.

Chan, T. F., Sandberg, B. Y., and Vese, L. A. (2000). Ac-

tive contours without edges for vector-valued images.

Journal of Visual Communication and Image Repre-

sentation, 11(2):130–141.

Cheng, I., Sun, X., Alsufyani, N., Xiong, Z., Major, P.,

and Basu, A. (2015). Ground truth delineation for

medical image segmentation based on local consis-

tency and distribution map analysis. In 37th Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC), pages 3073–

3076.

Claridge, E. and Orun, A. (2002). Modelling of edge pro-

ﬁles in pigmented skin lesions. In Proceedings of med-

ical image understanding and analysis, pages 53–56.

BIOIMAGING 2019 - 6th International Conference on Bioimaging

Do, T. T., Hoang, T., Pomponiu, V., Zhou, Y., Zhao, C.,

Cheung, N. M., Koh, D., Tan, A., and Hoon, T.

(2018). Accessible melanoma detection using smart-

phones and mobile image analysis. IEEE Transactions

on Multimedia, pages 1–1.

Ganster, H., Pinz, P., Rohrer, R., Wildling, E., Binder,

M., and Kittler, H. (2001). Automated melanoma

recognition. IEEE transactions on medical imaging,

20(3):233–239.

Garnavi, R., Aldeen, M., and Celebi, M. (2011). Weighted

performance index for objective evaluation of border

detection methods in dermoscopy images. Skin Re-

search and Technology, 17(1):35–44.

Hance, G. A., Umbaugh, S. E., Moss, R. H., and Stoecker,

W. V. (1996). Unsupervised color image segmenta-

tion: with application to skin tumor borders. IEEE

Engineering in Medicine and Biology Magazine,

15(1):104–111.

Haralick, R. M. and Shapiro, L. G. (1992). Computer and

robot vision. Addison-wesley.

Iyatomi, H., Oka, H., Celebi, M. E., Hashimoto, M., Hagi-

wara, M., Tanaka, M., and Ogawa, K. (2008). An

improved internet-based melanoma screening system

with dermatologist-like tumor area extraction algo-

rithm. Computerized Medical Imaging and Graphics,

32(7):566–579.

K´echichian, R., Gong, H., Revenu, M., Lezoray, O., and

Desvignes, M. (2014). New data model for graph-

cut segmentation: Application to automatic melanoma

delineation. In 2014 IEEE International Conference

on Image Processing (ICIP), pages 892–896.

Ker, J., Wang, L., Rao, J., and Lim, T. (2018). Deep learning

applications in medical image analysis. IEEE Access,

6:9375–9389.

Korotkov, K. and Garcia, R. (2012). Computerized anal-

ysis of pigmented skin lesions: A review. Artiﬁcial

intelligence in medicine, 56(2):69–90.

Lankton, S. and Tannenbaum, A. (2008). Localizing region-

based active contours. IEEE transactions on image

processing, 17(11):2029–2039.

Li, B. N., Chui, C. K., Chang, S., and Ong, S. H. (2011). In-

tegrating spatial fuzzy clustering with level set meth-

ods for automated medical image segmentation. Com-

puters in biology and medicine, 41(1):1–10.

Lloyd, S. (1982). Least squares quantization in pcm. IEEE

transactions on information theory, 28(2):129–137.

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.

Melli, R., Grana, C., and Cucchiara, R. (2006). Comparison

of color clustering algorithms for segmentation of der-

matological images. In Medical Imaging 2006: Image

Processing, volume 6144, page 61443S. International

Society for Optics and Photonics.

Mendonc¸a, T., Ferreira, P. M., Marques, J. S., Mar-

cal, A. R., and Rozeira, J. (2013). Ph2-a der-

moscopic image database for research and bench-

marking. In Engineering in Medicine and Biol-

ogy Society (EMBC), 2013 35th Annual International

Conference of the IEEE, pages 5437–5440. IEEE.

http://www.fc.up.pt/addi/ph2%20database.html.

Oliveira, R. B., Papa, J. P., Pereira, A. S., and Tavares,

J. M. R. (2018). Computational methods for pig-

mented skin lesion classiﬁcation in images: review

and future trends. Neural Computing and Applica-

tions, 29(3):613–636.

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

level histograms. IEEE transactions on systems, man,

and cybernetics, 9(1):62–66.

Pathan, S., Prabhu, K. G., and Siddalingaswamy, P. (2018).

Techniques and algorithms for computer aided diag-

nosis of pigmented skin lesions – a review. Biomedi-

cal Signal Processing and Control, 39:237–262.

Rajab, M., Woolfson, M., and Morgan, S. (2004). Applica-

tion of region-based segmentation and neural network

edge detection to skin lesions. Computerized Medical

Imaging and Graphics, 28(1):61–68.

Sahoo, P. K., Soltani, S., and Wong, A. K. (1988). A survey

of thresholding techniques. Computer vision, graph-

ics, and image processing, 41(2):233–260.

Sarkar, S., Paul, S., Burman, R., Das, S., and Chaudhuri,

S. S. (2014). A fuzzy entropy based multi-level im-

age thresholding using differential evolution. In In-

ternational Conference on Swarm, Evolutionary, and

Memetic Computing, pages 386–395. Springer.

Soille, P. (2013). Morphological image analysis: princi-

ples and applications. Springer Science & Business

Media.

Tasoulis, S., Doukas, C., Maglogiannis, I., and Plagianakos,

V. (2010). Classiﬁcation of dermatological images us-

ing advanced clustering techniques. In Engineering in

Medicine and Biology Society (EMBC), 2010 Annual

International Conference of the IEEE, pages 6721–

6724. IEEE.

Umbaugh, S. E., Moss, R. H., Stoecker, W. V., and

Hance, G. A. (1993). Automatic color segmenta-

tion algorithms-with application to skin tumor feature

identiﬁcation. IEEE Engineering in Medicine and Bi-

ology Magazine, 12(3):75–82.

Zhou, H., Chen, M., Zou, L., Gass, R., Ferris, L., Dro-

gowski, L., and Rehg, J. M. (2008). Spatially con-

strained segmentation of dermoscopy images. In

Biomedical Imaging: From Nano to Macro, 2008.

ISBI 2008. 5th IEEE International Symposium on,

pages 800–803. IEEE.

Image Segmentation using Gradient-based Histogram Thresholding for Skin Lesion Delineation