Hybrid channel and spatial attention-UNet for skin lesion segmentation

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 13, No. 1, March 2024, pp. 1077~1089
ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i1.pp1077-1089  1077
Journal homepage: http://ijai.iaescore.com
Hybrid channel and spatial attention-UNet for skin lesion
segmentation
Soumya Gadag1
, Pradeepa Palraj2
1
Department of Electronics and Communication, S. G. Balekundri Institute of Technology,
Research Scholar at Jain Deemed to be University, Bangalore, India
2
Department Electrical and Electronics Engineering, Faculty of Engineering and Technology,
Research Supervisor Jain Deemed to be University, Kanakapura Campus Bangalore, India
Article Info ABSTRACT
Article history:
Received May 12, 2023
Revised Oct 20, 2023
Accepted Nov 7, 2023
Melanoma is a type of skin cancer which has affected many lives globally.
The American Cancer Society research has suggested that it a serious type of
skin cancer and lead to mortality but it is almost 100% curable if it is
detected and treated in its early stages. Currently automated computer
vision-based schemes are widely adopted but these systems suffer from poor
segmentation accuracy. To overcome these issue, deep learning (DL) has
become the promising solution which performs extensive training for pattern
learning and provide better classification accuracy. However, skin lesion
segmentation is affected due to skin hair, unclear boundaries, pigmentation,
and mole. To overcome this issue, we adopt UNet based deep learning
scheme and incorporated attention mechanism which considers low level
statistics and high-level statistics combined with feedback and skip
connection module. This helps to obtain the robust features without
neglecting the channel information. Further, we use channel attention, spatial
attention modulation to achieve the final segmentation. The proposed DL
based scheme is instigated on publically available dataset and experimental
investigation shows that the proposed Hybrid Attention UNet approach
achieves average performance as 0.9715, 0.9962, 0.9710.
Keywords:
Computer vision
Deep learning
Image processing
Melanoma
Segmentation
This is an open access article under the CC BY-SA license.
Corresponding Author:
Soumya Gadag
Department of Electronics and Communication Engineering, S. G. Balekundri Institute of Technology
Jain Deemed to be University
Bangalore, Karnataka, India
Email: soumya.gadag5@gmail.com
1. INTRODUCTION
Cancer is considered as the most challenging issue for the modern medical science. Cancer is a
process under which body cells start growing unusually and becomes of control. This abnormal growth of
cells affects other stronger cells and deteriorates the body’s immune system [1]. These cancer cells may
spread to other areas of the body from existing body part. The liver, lung, bone, breast, prostate, bladder, and
skin are just a few of the bodily areas that can be impacted by cancer [1]. In this work, we emphasis on
studying the skin cancer and current advancements in this field to detect the skin cancer. Skin cancer is one
of the most dangerous cancers which may lead to death if not diagnosed early. It is an invasive disease
instigated due to the unconventional progression of melanocyte cells in the human body. Later, this abnormal
growth tends to reproduce and inflate through the lymph nodes resulting in destroying the surrounding tissues
[2]. These damaged cells lead to escalate a mole on the peripheral layer of skin, which is categorized as
benign, malignant. Generally, Melanoma (mel), Basal-cell carcinoma (BCC), non-melanoma skin cancer

 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 1, March 2024: 1077-1089
1078
(NMSC), and squamous-cell carcinoma (SCC) are the most widely known form of the skin cancer [3].
Melanoma is characterized by the malignancy of melanocytes which are known as the melanin (pigment)
producing cells in the basal layer of the epidermis. Below given Figure 1 illustrates the different categories of
skin cancer.
Figure 1 Skin cancer
According to a study presented by Majumder and Ullah [4] reported that around 287,723 melanoma
cases were identified globally and around 60,172 deaths were also reported. This study predicted total
number of expected melanoma cases 96,480 and 7230 deaths (including 2490 female and 4740 male) [5].
Similarly, International Agency for Research on Cancer (IARC) presented a study in [6] which reported the
19.3 million new cases and 10 million deaths globally by year of 2020. Moreover, 100,350 new cases were
identified in US and 6850 deaths also were reported in 2020. The American Cancer Society [7] projected that
in 2021 there would be 106,110 new instances of melanoma identified which includes around 62,260 and
43,850 men and women, and 7180 melanoma patients would die. Rahib et al. [8] predicted that by 2040 the
melanoma would overtake the lung cancer as second most common type of cancer. These studies have
reported the prevalence and severity of skin cancer. Therefore, early detection is considered as one of the
most beneficial solutions. Moreover, the early detection helps to increase survival rate up to 95% [9].
Currently, the advancement in digital image processing system has gained huge attention in various
real-time medical applications. Biomedical image analysis provides several critical information. Thus, recent
biomedical systems are widely dependent on the image analysis systems. In this field of skin cancer
detection, the dermatocsopy is considered as a promising solution because it provides magnified (typically by
x10) skin image. The manual analysis of these images requires more time and effort and accuracy also varies
according to the expert skills of the dermatologists. Therefore, researchers have focused on development of
automated process of biomedical image analysis for skin cancer detection. Several methods have been
presented to accomplish this task where ABCDE measurement method has been considered as a reliable
solution. This method considers several parameters such as asymmetry, border, colour, diameter over 6 mm,
and development. However, accuracy and reliability of this system depends on the skills of dermatologists.
Moreover, Dermatologists' ability to accurately detect skin lesions is reported to be less than 80% accurate
[10].
In order to detect the skin cancer, skin lesion segmentation is a crucial process where affected region
is extracted and used for further analysis for classification. Several segmentation methods have been
presented such as Hu et al. [11] presented adaptive thresholding approach by considering saliency maps in
wavelet domain, fuzzy logic [12], clustering [13], and region growing [14]. However, these methods suffer
from the poor accuracy issues moreover the complex nature of skin lesion images due to inaccurate
boundaries, colour variations affect the outcome of these segmentation methods. Currently, deep learning-
based methods have been adopted widely in bio medical image analysis. Ronneberger et al. [15] presented a
Convolutional Neural Network based architecture called as UNet which is based on the encoder and decoder
modules. This architecture uses down sampling, up sampling and skip connection modules to enhance the
segmentation outcome. Based on this architecture, several deep learning models have been presented for skin
lesion segmentation such as DSNet, separable-Unet and attention UNet. to improve the skin lesion
segmentation performance. Literature review presents a brief discussion about these methods.

Int J Artif Intell ISSN: 2252-8938 
Hybrid channel and spatial attention-UNet for skin lesion segmentation (Soumya Gadag)
1079
As discussed before, the main aim of skin lesion segmentation is to identify the lesion boundaries
from dermoscopic images and separating the lesion area from surrounding healthy skin or background.
Generally, skin lesion segmentation faces several challenges such as variability in Lesion appearance, image
quality, overlapped lesion, variations in anatomical location of body, and expertise of dermatologist. Several
methods have been developed such based on traditional approaches of image pre-processing methods
however, these methods face challenges due to variability and complexity of skin lesions. Similarly, machine
and deep learning-based methods also adopted for segmentation and classification using convolutional neural
networks (CNNs), Fully Convolutional Networks (FCNs), U-Net, and Mask R-CNN for semantic and
instance segmentation tasks. Though, computational complexity and classification accuracy remains
challenging issues in this domain. Therefore, development of efficient segmentation and classification
method is highly recommended to obtain the reliable classification performance. Rest of the article is
organized in following segments: section II presents literature review about various deep learning-based
segmentation methods, section III presents proposed deep learning based UNet architecture, section IV
presents outcome of proposed model and comparative analysis, finally, section V presents concluding
remarks about the research.
In this work, we focus on developing a new deep learning architecture for skin lesion segmentation
by using UNet as base architecture of work. The main contributions of this work are as:
a) To present a novel pre-processing method for hair removal to improve the image quality.
b) In next stage, we present a UNet based model along with the low and high order statistics for feature
extraction with attention mechanism.
c) The attention-based feature extraction mechanism also uses a feedback module to retain the rich
feature information.
d) Finally, we present channel and spatial attention module to obtain the final segmentation.
2. LITERATURE SURVEY
Previous section has described the importance of deep learning-based systems for skin cancer
segmentation and classification. This section presents the brief overview of existing schemes of skin lesion
and tumor segmentation. Conventional skin lesion segmentation procedures rely on thresholding, edge-
detection, active contours including level sets, clustering, expectation maximization and region-based
techniques such as region growing and region merging. However, the thresholding-based methods fail to
overcome the skin lesion complexities and artifacts. Similarly, the region and boundary-based methods
provide only a coarse segmentation. The threshold-based methods require a predefined threshold which may
not provide stable results due to noise and intensity variations of skin images. Clustering based approaches
require high-resolution images and requires addition pre-processing steps. Moreover, the performance of
these methods is influenced by the parameter tuning and finding the universal parameter is a challenging
task. Therefore, advanced research has focused on development of artificial intelligence-based approaches
[16].
Baig et al. [17] presented a detailed review about melanoma segmentation and classification.
Authors have reported that the skin lesion image analysis suffer from various issues such as poor contrast,
fuzzy borders, irregular boundaries, hairs, bubbles and various other kind of distortions. However, the current
advancement in deep learning schemes has led to increase the segmentation accuracy. Several studies have
reported that the semantic segmentation by using CNN led to providing the robust solution for segmentation
tasks by using pixel-wise classification.
Based on this concept of deep learning, Hasan et al. [18] presented an automated segmented
approach called as DSNet. Generally, the Deep Learning based methods require a greater number of training
parameters which increases the computational complexity and training duration. Therefore, authors presented
a lightweight approach which uses depth-wise separable convolutions to the replace the standard convolution.
The depth-wise convolution helps to discriminate the features on pixel space at different states of encoder.
UNet, an encoder –decoder based deep learning architecture has become popular in biomedical image
segmentation due to its property to use the global and local information at the same time [19]. Therefore,
several UNet based architectures are developed to increase the segmentation accuracy such as improved
UNet [20], separable-UNet [21], and U-Net++ [22]. According to the study presented in [20] the existing
UNet architecture is improved by incorporating the batch normalization and dilated convolution layers. The
dilated convolution layer helps to expand the perceptive field and batch normalization helps to mitigate the
overfitting problem without affecting the image resolution.
Tang et al. [21] presented separable-UNet which uses separable convolution blocks along with UNet
architectures. The separable convolutions use stochastic weight averaging that is useful in context feature
correlation and high-semantic feature information extraction. On the other hand, the stochastic weight
averaging method is used to handle the over-fitting issue. Zhao et al. [22] discussed the importance of fully

 ISSN: 2252-8938
1080
connected network and UNet. However, these networks are prone to gradient vanishing issue which increases
due to increase in depth. Thus, the reduced gradient result in the Jaccard index.
Generally, these deep learning models require huge computation resources to meet the
computational requirement of all activations. Recently, attention-based mechanisms are widely adopted in
this context which helps to highlight the most relevant activations during training process. Arora et al. [23]
adopted this mechanism and presented attention based deep learning mechanism by modifying the UNet
model which introduces Group Normalization in the encoder and decoder layers. This model also uses
attention gate in skip connection and further it is incorporated with Tversky Loss (TL). The group
normalization replaces Batch Normalization with GroupNorm to extract the feature maps. Moreover, the
attention module helps to identify the high-dimensional information from background region. Moreover, it
uses atrous convolutions to improve the feature reuse. Tong et al. [24] presented a modified UNet model by
incorporating the triple attention mechanism. The attention gate is used to compute the attention coefficients
which are used for selecting the regions. In next stage, a dual attention module is incorporated which has
spatial and channel attention blocks to estimate the spatial correlation among features. Dong et al. [25]
developed a feedback attention model which consist of feedback fusion block and attention block. The
amalgamation of these two blocks helps to achieve the rich feature mapping without employing data
enhancement.
Aghdam et al. [26] reported that the performance of traditional UNet is limited due to several
restrictions of convolution operations and capturing the long-term dependencies in UNet. To overcome this
issue, authors presented transformer based UNet architecture Attention Swin u-net where CNN blocks are
replaced with the Swin transformer blocks. This arrangement helps to capture the local and global
representation. The skip connection layer is useful in feature reusability and suggested to include the
attention mechanism to improve the performance of traditional skip connection. Jiang et al. [27] considered
UNet as baseline model and incorporated an attention mechanism to consider the position and context
information. Along with this, a global context information complementary module is also developed to
achieve the long-term dependencies. Similarly, the multi-scale grouped dilated convolution feature extraction
module (MSEM) is also presented to extract the multi-scale feature details.
This section has discussed about recent skin lesion segmentation methods where most of the recent
methods are based on the deep learning-based approaches. However, these methods face several challenges
such as unavailability of skin lesion dataset as limited datasets are available which leads to cause data
imbalance problem. Moreover, the current deep learning methods are trained on specific datasets which
doesn’t cover the diverse demographics skin types. Similarly, the supervised classification models require
precise annotations which becomes time consuming process. Some skin lesions, such as non-melanoma skin
cancers or lesions on anatomically complex regions, can be challenging to segment accurately. Research that
focuses on improving the segmentation of these challenging cases would be valuable.
3. PROPOSED MODEL
This section describes the proposed deep learning-based solution for skin lesion segmentation along
with pre-processing approach for hair removal. The hair removal process is based on the adaptive gradient
mechanism which helps to improve the boundary detection. Further, Dull Razor model is adopted for hair
removal [28]. Later, we present UNet Based deep learning architecture for segmentation. The proposed
architecture uses spatial and channel attention models to increase feature quality. Moreover, this model uses
small and large order statistics to extracts the features from different channels. Later, a feedback model is
used to obtain the information from decoder blocks.
3.1. Pre-processing for hair removal
This sub-section presents a brief discussion on proposed model for skin hair removal where a skin
lesion image is presented as 𝑢(𝑥), 𝒙 = (𝑥1, 𝑥2) ∈ 𝑹2
. This image is presented as local surface where (𝑖, 𝑗)
denotes the pixels position. The local shape of input image can be presented as shown in (1):
ℍ(𝒙, 𝜎) = (
𝜕2𝑤
𝜕𝑥1
2 (𝒙)
𝜕2𝑤
𝜕𝑥1𝜕𝑥2
(𝒙)
𝜕2𝑤
𝜕𝑥1𝜕𝑥2
(𝒙)
𝜕2𝑤
𝜕𝑥2
2 (𝒙)
) (1)
where 𝜎 denotes the standard deviation parameter and it is used for Gaussian Kernel representation for low-
pass filters, given as shown in (2):

1081
𝐺𝜎(𝑦) =
1
2𝜋𝜎2 exp (−
𝑦1
2+𝑦2
2
2𝜎2 ) (2)
where 𝒚 = (𝑦1, 𝑦2) and 𝑤 = 𝐺𝜎(𝑦) ∗ 𝑢 where ∗ denotes the 2D convolution operation. In this work, we use
symmetric Gaussian kernel. We compute the maximum and minimum Eigen values of ℍ to obtain the
principal curvatures as 𝜆𝑖𝑗
+
and 𝜆𝑖𝑗
−
, respectively. The obtained maximum curvature helps to identify the black
lines on images with bright background and the minimum curvature helps to obtain the light lines on dark
background. This method helps to overcome the issues of traditional gradient magnitude-based methods
because the existing methods detect only hair borders whereas the proposed model helps to produce the solid
structure. Further, we used Dull Razor model for hair removal process [29].
3.2. Hybrid spatial and channel attention UNet architecture
Traditional deep learning-based architectures are mainly focused on the developing the deep and
wider architecture for learning. However, these techniques ignore exploring the correlations among different
channels which is helpful in extracting the rich feature information. To address this issue, we introduce a
novel deep learning architecture which is based on the working of UNet segmentation model. In order to
increase the performance of segmentation, we incorporated a hybrid attention mechanism which considers
channel and spatial attention models to generate the rich feature information as shown in Figure 2 .
Figure 2. Channel and spatial attention modules
3.2.1. Hybrid model and its components
In traditional segmentation methods, the receptive field is minimized to local region. This restricts
the performance to obtain the broad and rich contextual information. Therefore, we introduce channel and
spatial attention mechanism. Below given figure depicts the architecture of channel attention module. This
map can be computed directed as 𝑋 ∈ 𝑅𝐶×𝐶
from actual features 𝐴 ∈ 𝑅𝐶×𝐻×𝑊
. Specifically, the original
features are reshaped as from 𝐴 to 𝑅𝐶×𝑁
and then perform product operation between 𝐴 and 𝐴′
(transpose of
A). Finally, we use SoftMax layer to achieve the channel attention map 𝑋 ∈ 𝑅𝐶×𝐶
as (3):
𝑥𝑖𝑗 =
exp(𝐴𝑖,𝐴𝑗)
∑ exp(𝐴𝑖,𝐴𝑗)
𝐶
𝑖=1
(3)
where 𝑥𝑖𝑗 represents the impact of 𝑖𝑡ℎ
channel on 𝑗𝑡ℎ
channel. Next, we perform multiplication between 𝑋′
and A. The obtained outcome is reshaped to 𝑅𝐶×𝐻×𝑊
. The obtained outcome is multiplied with scale
parameter 𝛽 and element –wise sum with 𝐴 to obtain the final outcome as (4):
𝐸𝑗 = 𝛽 ∑ (𝑥𝑗𝑖𝐴𝑖)
𝐶
𝑖=1 + 𝐴𝑗 (4)
where 𝛽 denotes the weight learning
Similarly, below given figure denotes the architecture of spatial attention mechanism. It consists of
lightweight squeeze and excitation block. Moreover, it uses a layer normalization before ReLU. This layer

 ISSN: 2252-8938
1082
normalization block increases the object detection and segmentation performance. This module can be
expressed as (5):
𝑍𝐼 = 𝐴𝑖 + 𝑊𝑣2𝑅𝑒𝐿𝑈 (𝐿𝑁(𝑊𝑣1 ∑
𝑒
𝑊𝑘𝐴𝑗
∑ 𝑒𝑊𝑘𝐴𝑚
𝑁𝑝
𝑚=1
𝐴𝑗))
𝑁𝑝
𝑗=1
(5)
where 𝐴 is the feature map ( also input), 𝑁𝑝 denotes the number of position in feature map, and A & Z
denotes the input and output, respectively,
𝑒
𝑊𝑘𝐴𝑗
∑ 𝑒𝑊𝑘𝐴𝑚
𝑁𝑝
𝑚=1
denotes the weight of global attention and 𝛿(. ) =
𝑊𝑣2𝑅𝑒𝐿𝑈(𝐿𝑁(𝑊𝑣1(. )) represents the bottleneck transformation. Finally, element-wise sum is performance
to obtain the final feature map.
The proposed Hybrid Spatial Channel Attention (HSCA) model uses smallest statistics and larger
order statistics to compute the correlation among attributes to obtain the power representation of attributes. It
contains two modules as small order statistic (SOSCA) and largest order statistic (LOSCA). The SOSCA
model uses global average pooling method to extract the small order features. Later, these features are
processed through a small customized Fully connected. The obtained outcome is further rescaled to achieve
the intermediate features. On the other hand, the LOSCA attention mechanism uses bilinear pooling to obtain
the intermediate features of different channels. Further, these features are used as input to Full connected
network. The final outcome of this hybrid model is obtained by performing the element-wise addition of
LOSCA and SOSCA. This helps to explore the correlation among features.
a. Small order channel attention block (SOCA)
In this stage, we use global average pooling to contract the gating attributes with the help of its
spatial dimensions. Let, 𝒳 is the attribute set denoted as: 𝒳 = [𝑋1, … , 𝑋𝑐, … 𝑋𝑐] ∈ ℝ𝐻×𝑊×𝐶
and its
corresponding gating features are given as 𝒴 = [𝑌1, … 𝑌𝑐, … . , 𝑌𝐶] ∈ ℝ𝐻×𝑊×𝐶
. During the max pooling
operation, it extracts the maximum value of each channel but it doesn’t not utilize the global information.
Therefore, we incorporate global average pooling operation which is expressed as expressed in Figure 3.
𝑦𝑐 =
1
𝐻×𝑊
∑ ∑ 𝑌𝑐(𝑖, 𝑗)
𝑊
𝑗=1
𝐻
𝑖=1 (6)
Figure 3. Proposed hybrid attention Unet
Later, these features are fed to the FC layers. Here, 𝜃1(𝑊1, 𝑏1) is the 𝐹𝐶1 and 𝜃2(𝑊2, 𝑏2) denotes the
𝐹𝐶2. The obtained channel attention weights denoted as 𝑓 = [𝑓1, … , 𝑓𝑐, … , 𝑓𝐶] ∈ 𝑅𝐶
which can be derived as
(7):
𝑓 = 𝜎(𝑊2
𝑇
(𝛿(𝑊1
𝑇
𝑦 + 𝑏1)) + 𝑏2) (7)
where 𝛿 characterise the rectified linear unit (ReLU) activation function and 𝜎 represents the sigmoid
activation function. The ReLU activation function helps to improve the non-linearity of the network whereas

1083
the Sigmoid function helps to scale the input as (0,1) to obtain the weights. This design helps to institute the
correlation among different channels to learn the intermediary features. With the help of attention weights,
the final features can be represented as (8):
𝑋
̂𝑐 = 𝑓𝑐. 𝑋𝑐 (8)
b. Largest order channel attention block (LOSCA)
The aforementioned attention module computes mean of the gating attributes. These gating
attributes are used in rescaling the input attributes. However, it fails to establish the correlation among
channel intermediate attributes. Therefore, we incorporate large order statistic computation. In this module
also, we consider the input as 𝒳 = [𝑋1, … , 𝑋𝑐, … 𝑋𝑐] ∈ ℝ𝐻×𝑊×𝐶
and its corresponding gating features are
given as 𝒴 = [𝑌1, … 𝑌𝑐, … . , 𝑌𝐶] ∈ ℝ𝐻×𝑊×𝐶
. According to this process, the bilinear pooling is applied on the
gating feature where 𝐵 represents the pooling matrix. this can be expressed as (9):
𝐵 =
[
𝑣𝑒𝑐(𝑌1)
…
𝑣𝑒𝑐(𝑌𝑐)
…
𝑣𝑒𝑐(𝑌𝐶)] [
𝑣𝑒𝑐(𝑌1)
…
𝑣𝑒𝑐(𝑌𝑐)
…
𝑣𝑒𝑐(𝑌𝐶)]
𝑇
∈ ℝ𝐶×𝐶
(9)
Here, 𝑣𝑒𝑐 denotes the vectorization operation, the obtained pooling matrix is semi-symmetric
positive definite in nature which contains inner-product of two different feature channels. This inner-product
is helpful in incorporating the high-order feature. Further, we apply, matrix logarithmic normalization model
to normalize the SPD matrix data. This matrix can be represented in the Eigen form as (10):
𝐵 = 𝑈∑𝑈𝑇
(10)
where 𝑈 denotes the Eigen-vector, ∑ = 𝑑𝑖𝑎𝑔(𝜎1, . . , 𝜎𝑐, . . 𝜎𝐶) denotes the diagonal matrix. the matrix
normalization can be represented as (11):
𝐵
̂ = 𝐵𝛼
= 𝑈∑𝛼
𝑈𝑇
∈ ℝ𝐶×𝐶
(11)
where 𝛼 represents the positive real number and ∑ =
𝛼
𝑑𝑖𝑎𝑔 (𝜎1
𝛼
, … , 𝜎𝑐
𝛼
, … . , 𝜎𝐶
𝛼
). With the help of this
normalized matrix 𝐵
̂ = [𝑏1, … , 𝑏𝑐, … , 𝑏𝐶] we obtain he channel wise attributes as 𝑑 ∈ ℝ𝐶
which is computed
as:
𝑑 =
1
𝐶
∑ 𝑏𝑐 ∈ ℝ𝐶
𝐶
𝑖 (12)
the obtained 𝑑 is processed through the fully connected networks given as 𝜃3(𝑊3, 𝑏3) and 𝜃4(𝑊4, 𝑏4). It
generates the attention weights 𝑠 ∈ ℝ𝐶
which is learned as (13).
𝑠 = 𝜃4(𝜃3(𝑑)) = 𝜎(𝑊4
𝑇)𝛿(𝑊3
𝑇
𝑑 + 𝑏3)) + 𝑏4) (13)
The outcome of SOCA is rescaled to generate the final output of SOCA 𝒳
̃𝑠, … 𝒳
̃𝑐, … , 𝒳
̃𝐶. Therefore, the final
outcome of this hybrid model is given as (14):
𝑋
̃𝑐 = 𝑠𝑐. 𝑋𝑐 (14)
3.2.2. Loss function
In this article, we use Binary Cross Entropy as the main objective function of training. Let us
consider that 𝑃𝑟𝑒𝑖 denotes the predicted output of 𝑖𝑡ℎ
model, ground truth of 𝑖𝑡ℎ
image is 𝐺𝑇𝑖, the pixel
values at 𝑖 and 𝑗 position in ground truth is 𝐺𝑇𝑖,𝑗 similarly, 𝑃𝑟𝑒𝑖,𝑗 denotes the pixel values at 𝑖 and 𝑗 positions
in predicted image. The binary cross entropy for these images can be expressed as (15):
𝐿𝐵𝐶𝐸 = 𝐵𝐶𝐸𝐿𝑜𝑠𝑠(𝑃𝑟𝑒𝑖, 𝐺𝑇𝑖)
= − ∑ ∑ [(𝐺𝑇𝑖,𝑗 × log 𝑃𝑟𝑒𝑖,𝑗) + (1 − 𝐺𝑇𝑖,𝑗) × log(1 − 𝑃𝑟𝑒𝑖,𝑗)]
𝐻
𝑗=1
𝑊
𝑖=1 (15)
where 𝑊 and 𝐻 denotes the width and height of the image, respectively.

 ISSN: 2252-8938
1084
4. RESULTS AND DISCUSSIONS
This section presents the detailed discussion on the outcome of proposed model. The proposed
model is implemented by using Python 3.8 running on windows platform where Keras and Tensor flow
model are used in backend. The performance of proposed approach is compared with several schemes in
terms of Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Sensitivity, and specificity for the
considered three different datasets.
4.1. Dataset details
In this work, we explore the diverse array of datasets critical to our analysis. We have measured the
performance of proposed approach for different datasets and compared the outcome with existing schemes.
PH2, International Skin Image Collaboration (ISIC) 2017, ISIC 2018 and Human Against Machine
(HAM)10000 dermoscopic images are considered for analysis and evaluation of the proposed model. This
dataset played a pivotal role in our segmentation tasks, allowing us to assess the accuracy of our models with
precision.
4.1.1. ISIC 2017 DATASET
The ISIC 2017 dataset is also known as International Skin Imaging Collaboration (ISIC)2017
Challenge dataset, is a comprehensive collection of dermoscopic images used in ISIC 2017 Skin lesion
analysis to Melanoma Detection challenge. This dataset contains total 2000 dermoscopic images which are
acquired with the of eliminating the surface reflection technique. The dataset includes images of various skin
lesion classes such as melanoma, seborrheic keratosis, and benign nevi(moles). The images in the ISIC 2017
dataset are of high quality, captured using dermoscopy, a non-invasive imaging technique. The ISIC 2017
dataset is publicly available and can be accessed through the official ISIC Archive.
4.1.2. ISIC 2018 DATASET
ISIC 2018 dataset plays a vital role in advancing research related to dermatology and computer-
aided diagnosis of skin lesions, particularly melanoma. There are 2594 images in this dataset, and each one
has a grand truth mask. This dataset is used for three different tasks such as lesion segmentation, lesion
feature analysis, and classification. The dataset has been divided into three sub-sections: test data, which
contains 520 images, evaluation data, which contains 259 images, and train data, which contains 1815
images. Additionally, we downsized the input photos from 2016×3024 pixels to 256×256 pixels in order to
lower the computational and network training costs.
4.1.3. PH2
The dataset includes images of different skin lesion classes, such as melanocytic lesions, nevi, and
seborrheic keratosis. These classes are important for training machine learning models to differentiate
between different types of skin lesions, including malignant and benign ones. This dataset contains 200
dermoscopic images of skin lesion. In this work, we have divided 100 images for training purpose and 100
images for evaluation purpose. PH2 contains high quality images of pigmented skin lesions frequently used
for evaluation and developing computer aided diagnostic systems for skin lesions classification and diagnosis
of diseases.
4.1.4. HAM 10000
HAM10000 is a public dataset of skin lesion images, which is designed to facilitate research on
automated skin lesion classification using machine learning and computer vision techniques. The dataset
consists of 10,015 dermatoscopic images of skin lesions, which are categorized into seven different classes
based on their clinical diagnosis. The dataset was released in 2018 by the International Skin Imaging
Collaboration (ISIC), which is a global network of experts in the field of skin imaging. The seven different
classes in the HAM10000 dataset are: Melanoma, Melanocytic nevus, Basal cell carcinoma, Actinic
keratosis, Benign keratosis-like lesions, Dermatofibroma, Vascular lesions
4.2. Evaluation metrics
In order to quantitatively assess the performance of proposed model, we used five different metrics
considered and popularly used to evaluate the segmentation performance. These metrics provide a
standardized way to measure the outcomes against the ground truth. These metrics are described as follows in
Table 1.
The skin lesion images capture the skin hair along with the affected region. This type of images
affects the diagnosis performance. Therefore, we apply image pre-processing technique for hair removal.
Below given

1085
Figure 4, shows some sample outcome of hair removal process. The obtained images are used for
further analysis where we measure the performance and compare with existing schemes. Table 2 shows the
comparative performance analysis for ISIC 2017 dataset.
Table 1. Evalution metrics
Matrix Formula Description
Dice similarity
coefficient 𝐷𝑆𝐶(𝑆, 𝐺) =
2 × |𝑆⋂𝐺|
|𝑆| + |𝐺|
It is used to compute the spatial overall between predicted
and groundtruth image.
Intersection over
Union 𝐼𝑜𝑈(𝑆, 𝐺) =
|𝑆⋂𝐺|
|𝑆 ∪ 𝐺|
It is used to evaluate the segmentation accuracy. It is
measured by evaluating the ratio of object and
corresponding union which are projected in the same
plane
F-Measure
𝐹𝑀𝑒𝑎𝑠𝑢𝑟𝑒 =
2 × 𝑃 × 𝑅
𝑃 + 𝑅
It denotes the weighted harmonic mean of precision and
recall
Sensitivity
𝑆𝑒𝑛(𝑆, 𝐺) =
|𝑆 ∩ 𝐺|
|𝐺|
It measures the percentage of correctly classified pixels
Specificity
𝑆𝑝𝑒(𝑆, 𝐺) =
|(1 − 𝑆) ∩ (1 − 𝐺)|
|1 − 𝐺|
It measures the correct classification percentage of pixels
of background class
Figure 4. Outcome of hair removal process
Table 2. Comparative analysis for ISIC 2017 dataset
Technique Accuracy Sensitivity Specificity F1-score Jaccard Similarity
Full resolution CNN [30] 0.934 0.825 0.975 0.884 0.9365
R2UNet [30] 0.9424 0.9414 0.9425 0.8920 0.9421
MCGUNet [31] 0.9570 0.8502 0.9855 0.8927 0.9570
Baseline [32] 0.9314 0.9479 0.9263 0.8682 0.9314
Multiscale UNet [32] 0.9576 0.8870 0.9714 0.9032 0.9576
Proposed Model 0.9762 0.9516 0.9899 0.9215 0.9718
The proposed approach achieves the average performance as 0.9762, 0.9516, 0.9899, 0.9215, and
0.9718 in terms of Accuracy, Sensitivity, Specificity, F1-score, and Jaccard Similarity for ISIC 2017 dataset.
The proposed approach outperforms the recent attention and UNet based segmentation models. Similarly, we
measure the performance for ISIC 2018 dataset. Table 3 shows the comparative analysis for ISIC 2018
dataset.
The proposed approach obtained highest performance as 0.968,0.99, 0.999, 0.912, and 0.977 in
terms of Accuracy, Sensitivity, Specificity, F1-score, and Jaccard Similarity, respectively. The studies
presented in [30] and [33] reported the importance of Attention mechanism thereby we have adopted the
attention process and improved its performance. Further, we evaluate the performance of proposed approach
for PH2
dataset. The obtained performance is presented in Table 4. Further, the performance of proposed

 ISSN: 2252-8938
1086
approach is compared with several schemes in terms of DSC, IoU, Sensitivity, and specificity for the
considered three different datasets. Table 5 shows the comparative analysis with state-of-art algorithms.
Table 3. Comparative analysis for ISIC 2018 dataset
Technique Accuracy Sensitivity Specificity F1-score Jaccard similarity
R2UNet [30] 0.880 0.792 0.928 0.679 0.581
Attention R2UNet [30] 0.904 0.726 0.971 0.691 0.592
MCGUNet [31] 0.848 0.986 0.895 0.955
Baseline [32] 0.890 0.708 0.964 0.647 0.549
Multiscale UNet [32] 0.949 0.841 0.979 0.896 0.956
Attention UNet[33] 0.897 0.717 0.967 0.665 0.566
Proposed Model 0.968 0.99 0.999 0.912 0.977
Table 4. Comparative analysis for PH2
dataset
Technique Accuracy Sensitivity Specificity DICE Jaccard Similarity
Full Convolution Network [23] 0.9282 0.903 0.9402 0.8903 0.8022
FrCN [30] 0.9508 0.9372 0.9565 0.9177 0.8479
MCGUNet [31] 0.9537 0.8322 0.9714 0.9263 0.9537
Baseline [32] 0.9255 0.8163 0.9776 0.8761 0.7795
Multiscale UNet [32]
SegNet[34]
0.9617
0.9336
0.943
0.8653
0.9698
0.966
0.9377
0.8936
0.9617
0.8077
Proposed Model 0.9715 0.9962 0.9710 0.9997 0.9815
Table 5. Comparative analysis for different dataset
Dataset Techniques mDSC mIoU mSen mSpe
PH2 UNet [16] 0.876 0.780 0.816 0.978
Al et al. [29] 0.919 0.849 0.935 0.956
SegNet [34] 0.894 0.808 0.865 0.966
Double UNet [34] 0.907 0.899 0.945 0.966
Goyal et al [35] 0.910 0.838 0.933 0.930
Hasan et al. [18] - 0.870 0.929 0.968
Ozturk et al. [36] 0.931 0.872 0.968 0.954
Xie et al. [37] 0.920 0.858 0.964 0.945
Unver et al. [38] 0.882 0.798 0.837 0.941
Bi et al. [39] 0.907 0.840 0.948 0.942
Bi et al. [40] 0.921 0.859 0.965 0.946
MFSNet [41] 0.954 0.914 0.995 0.997
Proposed 0.9610 0.935 0.996 0.997
ISIC 2017 UNet [17] 0.778 0.683 0.812 0.805
SegNet [34] 0.821 0.696 0.801 0.954
Double UNet [35] 0.913 0.918 0.963 0.974
Ozturk et al. [36] 0.793 0.872 0.898 0.951
Xie et al. [37] - 0.776 0.879 0.956
Bi et al. [40] 0.844 0.749 0.910 0.925
MFSNet [41] 0.988 0.975 0.99 0.999
Al et al. [29] 0.855 0.773 0.825 0.980
Proposed 0.991 0.981 0.991 0.999
HAM10000 UNet [16] 0.781 0.774 0.799 0.802
Double UNet [34] 0.843 0.812 0.961 0.845
SegNet [34] 0.816 0.821 0.867 0.854
MFSNet [34] 0.906 0.902 0.999 0.99
Saha et al [42] 0.891 0.819 0.824 0.981
Abraham et al. [44] 0.856 - - -
Shahin et al [45] 0903 0.837 0.902 0.974
Bissoto et al [46] 0.873 0.792 0.934 0.936
Ibtehaz et al. [47] - 0.803 - -
Proposed 0.951 0.943 0.9510 0.9810
The experimental analysis shows that the proposed approach achieves better performance when
compared with existing schemes. The mean Dice score performance is obtained as 0.9610, 0.991 and 0.951
for PH2, ISIC2017 and HAM1000, respectively. Figure 5 demonstrates the qualitative results for PH2, ISIC
2017 and HAM dataset. The SegNet model suffer from the blurriness issue which is further resolved by
employing UNet however contextual information issue still remains challenging issue. In order to resolve this
issue DoubleUNet architecture is developed which is able to resolve this issue but fail to handle the complex
boundaries. Therefore, we presented a attention based mechanism which covers these issues and improves
the segmentation performance which is validated by quantitative and qualitative analysis.

1087
Sample results for PH2 dataset
Groundtruth SegNet UNet Double UNet MFSNet Proposed
Sample results for ISIC 2017 dataset
Sample results for HAM 10000 dataset
Figure 5. Sample outcome for PH2, ISIC 2017 and HAM 10000 dataset
5. CONCLUSION
In this work, we have focused on improving the segmentation accuracy of melanoma skin cancer.
Traditional methods are based on the computer vison-based tasks however the accuracy of these systems is
highly dependent on the expert skills. Deep learning-based schemes have gained huge attention recently.
Therefore, we present UNet based deep learning architecture which uses low and high statistics attention
information with feedback module. Moreover, the proposed approach uses channel and spatial attention
blocks to improve the channel attributes. The proposed approach is tested on publically available dataset and
extensive experimental study is presented which shows that the proposed approach realizes average
performance as 0.991, 0.981, 0.991, and 0.999 in terms of DSC, IoU, Sensitivity and Specificity.
REFERENCES
[1] D. N. H. Thanh, V. B. S. Prasath, L. M. Hieu, and N. N. Hien, “Melanoma skin cancer detection method based on adaptive
principal curvature, colour normalisation and feature extraction with the ABCD rule,” Journal of digital imaging, vol. 33, no. 3,
pp. 574–585, Jun. 2020, doi: 10.1007/s10278-019-00316-x.
[2] Y. N. Fu’adah, N. K. C. Pratiwi, M. A. Pramudito, and N. Ibrahim, “Convolutional neural network (CNN) for automatic skin
cancer classification system,” IOP Conference Series: Materials Science and Engineering, vol. 982, no. 1, p. 12005, 2020, doi:
10.1088/1757-899x/982/1/012005.
[3] D. Crisan, S. Kastler, K. Scharffetter‐Kochanek, M. Crisan, and L. Schneider, “Ultrasonographic assessment of depth infiltration
in melanoma and non‐melanoma skin cancer,” Journal of Ultrasound in Medicine, vol. 42, no. 7, pp. 1609–1616, 2023, doi:
10.1002/jum.16180.
[4] S. Majumder and M. A. Ullah, “Feature extraction from dermoscopy images for an effective diagnosis of melanoma skin cancer,”
2018 10th International Conference on Electrical and Computer Engineering (ICECE). IEEE, 2018. doi:

 ISSN: 2252-8938
1088
10.1109/icece.2018.8636712.
[5] American Cancer Society, “Melanoma skin cancer.” Accessed: Apr. 25, 2023. [Online]. Available:
https://www.cancer.org/cancer/melanoma-skin-cancer.html
[6] International Agency for Research on Cancer, “Cancer topics.” Accessed: Apr. 25, 2023. [Online]. Available:
https://www.iarc.who.int/cancer-topics/
[7] American Cancer Society, “Key statistics for melanoma skin cancer.” Accessed: Apr. 07, 2023. [Online]. Available:
https://www.cancer.org/cancer/types/melanoma-skin-cancer/about/key-statistics.html
[8] L. Rahib, M. R. Wehner, L. M. Matrisian, and K. T. Nead, “Estimated projection of US cancer incidence and death to 2040,”
JAMA network open, vol. 4, no. 4, pp. e214708–e214708, Apr. 2021, doi: 10.1001/jamanetworkopen.2021.4708.
[9] K. D. Miller et al., “Cancer treatment and survivorship statistics, 2016,” CA: A Cancer Journal for Clinicians, vol. 66, no. 4, pp.
271–289, 2016, doi: 10.3322/caac.21349.
[10] R. Kaur, H. GholamHosseini, R. Sinha, and M. Lindén, “Melanoma classification using a novel deep convolutional neural
network with dermoscopic images,” Sensors (Basel, Switzerland), vol. 22, no. 3, p. 1134, Feb. 2022, doi: 10.3390/s22031134.
[11] K. Hu et al., “A skin lesion segmentation method based on saliency and adaptive thresholding in wavelet domain,” Cognitive
Internet of Things: Frameworks, Tools and Applications. Springer International Publishing, pp. 445–453, 2019. doi: 10.1007/978-
3-030-04946-1_43.
[12] J. L. Garcia-Arroyo and B. Garcia-Zapirain, “Segmentation of skin lesions in dermoscopy images using fuzzy classification of
pixels and histogram thresholding,” Computer Methods and Programs in Biomedicine, vol. 168, pp. 11–19, 2019, doi:
10.1016/j.cmpb.2018.11.001.
[13] B. S. Lin, K. Michael, S. Kalra, and H. R. Tizhoosh, “Skin lesion segmentation: U-Nets versus clustering,” 2017 IEEE
Symposium Series on Computational Intelligence (SSCI). IEEE, 2017. doi: 10.1109/ssci.2017.8280804.
[14] Y. Guo, A. Ashour, and F. Smarandache, “A novel skin lesion detection approach using neutrosophic clustering and adaptive
region growing in dermoscopy images,” Symmetry, vol. 10, no. 4, p. 119, 2018, doi: 10.3390/sym10040119.
[15] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” Lecture Notes
in Computer Science. Springer International Publishing, pp. 234–241, 2015. doi: 10.1007/978-3-319-24574-4_28.
[16] Y. Ren, L. Yu, S. Tian, J. Cheng, Z. Guo, and Y. Zhang, “Serial attention network for skin lesion segmentation,” Journal of
Ambient Intelligence and Humanized Computing, vol. 13, no. 2, pp. 799–810, 2021, doi: 10.1007/s12652-021-02933-3.
[17] R. Baig, M. Bibi, A. Hamid, S. Kausar, and S. Khalid, “Deep learning approaches towards skin lesion segmentation and
classification from dermoscopic images - A review,” Current Medical Imaging Formerly Current Medical Imaging Reviews, vol.
16, no. 5, pp. 513–533, 2020, doi: 10.2174/1573405615666190129120449.
[18] M. K. Hasan, L. Dahal, P. N. Samarakoon, F. I. Tushar, and R. Martí, “DSNet: Automatic dermoscopic skin lesion segmentation,”
Computers in Biology and Medicine, vol. 120, p. 103738, 2020, doi: 10.1016/j.compbiomed.2020.103738.
[19] S. N. Hasan, M. Gezer, R. A. Azeez, and S. Gulsecen, “Skin lesion segmentation by using deep learning techniques,” 2019
Medical Technologies Congress (TIPTEKNO). IEEE, 2019. doi: 10.1109/tiptekno.2019.8895078.
[20] L. Liu, L. Mou, X. X. Zhu, and M. Mandal, “Skin lesion segmentation based on improved u-net,” 2019 IEEE Canadian
Conference of Electrical and Computer Engineering (CCECE). IEEE, 2019. doi: 10.1109/ccece.2019.8861848.
[21] P. Tang et al., “Efficient skin lesion segmentation using separable-Unet with stochastic weight averaging,” Computer Methods
and Programs in Biomedicine, vol. 178, pp. 289–301, 2019, doi: 10.1016/j.cmpb.2019.07.005.
[22] C. Zhao, R. Shuai, L. Ma, W. Liu, and M. Wu, “Segmentation of skin lesions image based on U-Net + +,” Multimedia Tools and
Applications, vol. 81, no. 6, pp. 8691–8717, 2022, doi: 10.1007/s11042-022-12067-z.
[23] R. Arora, B. Raman, K. Nayyar, and R. Awasthi, “Automated skin lesion segmentation using attention-based deep convolutional
neural network,” Biomedical Signal Processing and Control, vol. 65, p. 102358, 2021, doi: 10.1016/j.bspc.2020.102358.
[24] X. Tong, J. Wei, B. Sun, S. Su, Z. Zuo, and P. Wu, “ASCU-Net: Attention gate, spatial and channel attention U-net for skin lesion
segmentation,” Diagnostics (Basel, Switzerland), vol. 11, no. 3, p. 501, Mar. 2021, doi: 10.3390/diagnostics11030501.
[25] Y. Dong, L. Wang, S. Cheng, and Y. Li, “FAC-Net: Feedback attention network based on context encoder network for skin lesion
segmentation,” Sensors (Basel, Switzerland), vol. 21, no. 15, p. 5172, Jul. 2021, doi: 10.3390/s21155172.
[26] E. K. Aghdam, R. Azad, M. Zarvani, and D. Merhof, “Attention Swin U-Net: Cross-Contextual Attention Mechanism for Skin
Lesion Segmentation,” in 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), IEEE, Apr. 2023, pp. 1–5.
doi: 10.1109/ISBI53787.2023.10230337.
[27] Y. Jiang, J. Dong, Y. Zhang, T. Cheng, X. Lin, and J. Liang, “PCF-Net: Position and context information fusion attention
convolutional neural network for skin lesion segmentation,” Heliyon, vol. 9, no. 3, pp. e13942–e13942, Feb. 2023, doi:
10.1016/j.heliyon.2023.e13942.
[28] T. Lee, V. Ng, R. Gallagher, A. Coldman, and D. McLean, “Dullrazor®: A software approach to hair removal from images,”
Computers in Biology and Medicine, vol. 27, no. 6, pp. 533–543, 1997, doi: 10.1016/s0010-4825(97)00020-6.
[29] M. A. Al-masni, M. A. Al-antari, M.-T. Choi, S.-M. Han, and T.-S. Kim, “Skin lesion segmentation in dermoscopy images via
deep full resolution convolutional networks,” Computer Methods and Programs in Biomedicine, vol. 162, pp. 221–231, 2018, doi:
10.1016/j.cmpb.2018.05.027.
[30] M. Z. Alom, C. Yakopcic, T. M. Taha, and V. K. Asari, “Nuclei segmentation with recurrent residual convolutional neural
networks based U-Net (R2U-Net),” NAECON 2018 - IEEE National Aerospace and Electronics Conference. IEEE, 2018. doi:
10.1109/naecon.2018.8556686.
[31] R. Azad, M. Asadi-Aghbolaghi, M. Fathy, and S. Escalera, “Attention deeplabv3+: Multi-level context attention mechanism for
skin lesion segmentation,” Computer Vision – ECCV 2020 Workshops. Springer International Publishing, pp. 251–266, 2020. doi:
10.1007/978-3-030-66415-2_16.
[32] M. D. Alahmadi, “Multiscale Attention U-Net for skin lesion segmentation,” IEEE Access, vol. 10, pp. 59145–59154, 2022, doi:
10.1109/access.2022.3179390.
[33] O. Oktay et al., “Attention U-Net: Learning where to look for the pancreas,” ArXiv.org, Apr. 2018, doi:
10.48550/arXiv.1804.03999.
[34] D. Jha, M. A. Riegler, D. Johansen, P. Halvorsen, and H. D. Johansen, “DoubleU-Net: A deep convolutional neural network for
medical image segmentation,” 2020 IEEE 33rd International Symposium on Computer-Based Medical Systems (CBMS). IEEE,
2020. doi: 10.1109/cbms49503.2020.00111.
[35] M. Goyal, A. Oakley, P. Bansal, D. Dancey, and M. H. Yap, “Skin lesion segmentation in dermoscopic images with ensemble
deep learning methods,” IEEE Access, vol. 8, pp. 4171–4181, 2020, doi: 10.1109/access.2019.2960504.
[36] Ş. Öztürk and U. Özkaya, “Skin lesion segmentation with improved convolutional neural network,” Journal of digital imaging,
vol. 33, no. 4, pp. 958–970, Aug. 2020, doi: 10.1007/s10278-020-00343-z.

1089
[37] F. Xie, J. Yang, J. Liu, Z. Jiang, Y. Zheng, and Y. Wang, “Skin lesion segmentation using high-resolution convolutional neural
network,” Computer Methods and Programs in Biomedicine, vol. 186, p. 105241, 2020, doi: 10.1016/j.cmpb.2019.105241.
[38] H. M. Ünver and E. Ayan, “Skin lesion segmentation in dermoscopic images with combination of YOLO and GrabCut
algorithm,” Diagnostics (Basel, Switzerland), vol. 9, no. 3, p. 72, Jul. 2019, doi: 10.3390/diagnostics9030072.
[39] L. Bi, J. Kim, E. Ahn, A. Kumar, M. Fulham, and D. Feng, “Dermoscopic image segmentation via multistage fully convolutional
networks,” IEEE Transactions on Biomedical Engineering, vol. 64, no. 9, pp. 2065–2074, 2017, doi:
10.1109/tbme.2017.2712771.
[40] L. Bi, J. Kim, E. Ahn, A. Kumar, D. Feng, and M. Fulham, “Step-wise integration of deep class-specific learning for dermoscopic
image segmentation,” Pattern Recognition, vol. 85, pp. 78–89, 2019, doi: 10.1016/j.patcog.2018.08.001.
[41] H. Basak, R. Kundu, and R. Sarkar, “MFSNet: A multi focus segmentation network for skin lesion segmentation,” Pattern
Recognition, vol. 128, p. 108673, 2022, doi: 10.1016/j.patcog.2022.108673.
[42] A. Saha, P. Prasad, and A. Thabit, “Leveraging adaptive color augmentation in convolutional neural networks for deep skin lesion
segmentation,” 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI). IEEE, 2020. doi:
10.1109/isbi45749.2020.9098344.
[43] V. Badrinarayanan, A. Kendall, and R. Cipolla, “SegNet: A deep convolutional encoder-decoder architecture for image
segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 12, pp. 2481–2495, 2017, doi:
10.1109/tpami.2016.2644615.
[44] N. Abraham and N. M. Khan, “A novel focal Tversky loss function with improved attention u-net for lesion segmentation,” 2019
IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). IEEE, 2019. doi: 10.1109/isbi.2019.8759329.
[45] A. H. Shahin, K. Amer, and M. A. Elattar, “Deep convolutional encoder-decoders with aggregated multi-resolution skip
connections for skin lesion segmentation,” 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). IEEE,
2019. doi: 10.1109/isbi.2019.8759172.
[46] A. Bissoto, F. Perez, V. Ribeiro, M. Fornaciali, S. Avila, and E. Valle, “Deep-learning ensembles for skin-lesion segmentation,
analysis, classification: Recod titans at ISIC challenge 2018,” ArXiv.org, Aug. 2018, [Online]. Available:
http://arxiv.org/abs/1808.08480
[47] N. Ibtehaz and M. S. Rahman, “MultiResUNet : Rethinking the U-Net architecture for multimodal biomedical image
segmentation,” Neural Networks, vol. 121, pp. 74–87, 2020, doi: 10.1016/j.neunet.2019.08.025.
BIOGRAPHIES OF AUTHORS
Soumya Gadag completed her B.E degree in Electronics and Communication
Engineering from Visvesvaraya Technological University, India in 2008. She received her
M.Tech. from Visvesvaraya Technological University in 2012.Currently working as
Assistant Professor in the Department of Electronics and Communication Engineering, S.G.
Balekundri Institute of Technology, India. Her research interests include wireless
communication, image processing, machine learning and artificial intelligence. She can be
contacted at email: soumya.gadag5@gmail.com.
Dr. Pradeepa Palraj received the B.E. degree in Electrical and Electronics
Engineering Engineering from Government College of Technology, Coimbatore in 2001
and pursued her masters in Applied Electronics from Government College of Engineering,
Salem in 2004. Ph.D. degree from Anna University, Chennai in the year 2017. Currently
she is Professor and Head of the Electrical and Electronics Engineering Department.
Faculty of Engineering and Technology. Jain Deemed to be University. Her research
interests are image processing registration and segmentation of medical images. She can be
contacted at email: p.pradeepa@jainuniversity.ac.in.

Hybrid channel and spatial attention-UNet for skin lesion segmentation

More Related Content

Similar to Hybrid channel and spatial attention-UNet for skin lesion segmentation (20)

More from IAESIJAI (20)

Recently uploaded (20)

Hybrid channel and spatial attention-UNet for skin lesion segmentation