Transfer-learning based skin cancer diagnosis using fine-tuned AlexNet by marine predators algorithm

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 13, No. 4, December 2024, pp. 4822~4832
ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i4.pp4822-4832  4822
Journal homepage: http://ijai.iaescore.com
Transfer-learning based skin cancer diagnosis using fine-tuned
AlexNet by marine predators algorithm
Maha Ibrahim Khaleel1,2
, Amir Lakizadeh1
1
Department of Computer Engineering and Information Technology, University of Qom, Qom, Iran
2
Department of Computer Techniques Engineering, Alsafwa University College, Kerbala, Iraq
Article Info ABSTRACT
Article history:
Received Jan 16, 2024
Revised Mar 3, 2024
Accepted Mar 21, 2024
Melanoma represents one of the most dangerous manifestations of skin
cancer. According to statistics, 55% of patients with skin cancer have lost their
lives as a result of this disease. Early diagnosis of this condition will
significantly reduce mortality rates and save lives. In recent years, deep
learning methods have shown promising results and captured the attention of
researchers in this field. One common approach is the use of pre-trained deep
neural networks. In this work, a pre-trained AlexNet networks, which are
networks with specified architecture and weights is used to automatic skin
melanoma diagnosis. In the transfer learning phase, by reducing the learning
rate, the pre-trained network is trained to recognize skin cancer, which is
called fine-tuning. In addition, hyperparameters of the AlexNet network have
been optimized by the marine predators algorithm (MPA) algorithm to
enhance the network performance. Experimental findings show the
satisfactory efficiency of the presented approach, with an accuracy rate of
98.47%. The outcomes demonstrate the effectiveness of the suggested
approach in contrast to alternative existing methods.
Keywords:
AlexNet
Convolutional neural network
Marine predators algorithm
Skin cancer
Transfer learning
This is an open access article under the CC BY-SA license.
Corresponding Author:
Maha Ibrahim
Department of Computer Engineering and Information Technology, University of Qom
Qom, Iran
Email: m.ibrahim@stu.qom.ac.ir
1. INTRODUCTION
The human body comprises various organs, with one of the most prominent being the skin, which
serves as the body's largest organ, encompassing its entirety [1]–[5]. A skin ailment pertains to any condition
that impacts the human skin [6]–[9]. Skin diseases, including skin cancer, are regarded as among the most
widespread contagious conditions globally. Skin cancer, specifically, is a prevalent type of cancer that impacts
numerous individuals globally. It is identified by the abnormal proliferation of cells. Early detection is crucial
for effective treatment, as late-stage skin cancer can spread to other organs and potentially lead to death.
Identifying skin cancer in its initial phases is typically more successful. In the past, skin cancer diagnosis
involved using a dermo scope, which was expensive and required the expertise of a trained dermatologist. Skin
diseases can be caused by viruses, bacteria, allergies, fungal infections, and genetic factors. Typically, these
illnesses target the epidermis, the top layer of the skin, and their visibility can lead to psychological distress
and physical injuries [10]–[15]. Different varieties of skin lesions are present, including actinic keratosis (AK),
basal cell carcinoma (BCC), benign keratosis (BKL), dermatofibroma (DF), melanoma (MEL), melanocytic
nevus (NV), squamous cell carcinoma (SCC), and vascular lesion (VASC). The symptoms and severity of
these lesions vary, with certain ones being permanent while others are temporary. They can also vary in terms
of pain levels. Melanoma is regarded as the most perilous among these skin conditions and potentially deadly.

Int J Artif Intell ISSN: 2252-8938 
Transfer-learning based skin cancer diagnosis using fine-tuned AlexNet by marine … (Maha Ibrahim)
4823
Detecting skin diseases early is crucial, as around 95% of patients can recover if the condition is identified in
its initial stages. Leveraging an automated computer-aided system can be advantageous in precisely
categorizing skin diseases [11], [16]–[20]. There is frequently a considerable disparity between dermatologists
and skin disease patients since many individuals lack awareness of the various kinds, symptoms, and phases
of skin diseases. Delayed onset of symptoms can further complicate the situation, emphasizing the importance
of early detection. However, accurately diagnosing skin diseases to identify their type and stage can be
challenging and costly. Fortunately, the development of automatic computer-aided systems utilizing machine
learning techniques have enabled this possibility to achieve more accurate and rapid detection of skin disease
types. This advancement has the potential to bridge the gap and improve outcomes for patients.
Over the past 30 years, skin disease classification has been a significant area of research and has
become a popular topic. Despite the considerable effort put into researching skin disease detection and
classification, there remains an existing gap that requires attention and resolution. Past research endeavors have
predominantly concentrated on a single disease., indicating a need for additional research and development to
enhance the precision and utility of skin disease classification systems across a broader range of skin diseases
[21]–[24]. The existing research in this field is insufficient for effectively classifying multiple classes of skin
diseases. The task of classifying multiple classes is particularly challenging due to the similarities in behavior
exhibited by different skin diseases. With the advancement of computational technology, particularly machine
learning and computer vision, disease classification has improved. Imaging technologies are beneficial due to
their lower cost, ease of use, and non-invasiveness procedure. When machine learning and computer vision are
combined, the classification of skin lesions and selected features significantly impacts classification results.
convolutional neural networks (CNN), a recently technology based on deep learning, enable image
classification without the need for human detection and feature segmentation. This paper introduces an
innovative approach to skin cancer diagnosis leveraging an improved AlexNet network enhanced by the marine
predators algorithm (MPA). The research aims to increase the accuracy of the proposed method in comparison
with other past studies in the field of skin cancer detection. In summary, the primary contributions of our
research are outlined below:
− In this work, we have used pre-trained AlexNet networks, which are networks with specified architecture
and weights. In the transfer learning phase, by reducing the learning rate, we train the pre-trained network
to recognize skin cancer, which is called fine-tuning.
− The advantages of the fine-tuning method used in this work is the high learning ability on limited input
images, as well as the ability to decrease the diagnosis error.
In the proposed method, the MPA is used to optimally adjust the hyperparameters of the model, which
prevents overfitting of the network. The subsequent sections of this study are organized in the following
manner: section 2 delves into the relevant literature. The suggested method is denoted in section 3. Section 4
provides the datasets utilized in this study, along with the corresponding experimental results. In conclusion,
section 5 summarizes the research findings and outlines future prospects.
2. RELATED WORKS
Dorj et al. [2] employed dermoscopy pictures and digital pictures to distinguish skin disorders.
Authors utilized CNN in feature extraction phase, wherein support vector machine (SVM) is employed as the
classification method. It should be noted that in order to accurate skin diseases recognition from dermoscopy
pictures, the expertise of a dermatologist is required. The authors employed Gaussian channels for hair removal
and segmentation to isolate the affected areas. SVM was then utilized to classify the different types of skin
diseases. However, additional investigation is required to expand and enhance the skin diseases classification
specifically from dermoscopy pictures. Hosny et al. [25] suggested technique underwent assessment utilizing
a dataset called HAM10000. Authors attained enhanced test and training accuracy through the using of SVM
algorithm. However, analyzing the images posed challenges due to problematic elements such as reflections of
light from the skin surface and variations within the images.
The analysis of skin lesions model proposed in [13] focuses on the automated image analysis module,
that comprises stages including: image capturing, hair detection and elimination, lesion delineation, feature
derivation, and characterization. However, it is important to note that this framework specifically focuses on
identifying a single type of skin cancer and does not aim to distinguish between different types of skin tumors.
Shanthi et al. [26] introduced a model based on computer vision for diagnosing four main skin ailments. Their
methodology involved employing CNN networks with eleven layers, encompassing activation, convolutional,
fully connected, pooling, and soft-max layers. The evaluation of the model utilized images sourced from the
DermNet database, covering a range of skin disorders. However, the authors concentrated solely on four class
of skin disorders: urticaria, eczema herpeticum, keratosis, and acne, with a restricted number of samples
(30 to 60 samples per class). This research primary constraints entail the limited number of images and the
narrow focus on only four classes of skin diseases.

 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 4, December 2024: 4822-4832
4824
Bhavani et al. [27] proposed an approach based on computer vision to detect various dermatological
skin disorders. They utilized 3 deep learning-based methods, namely Resnet, Mobilenet, and Inception v3 for
feature extraction from medical images. Logistic regression, a ML technique, was utilized for training and
evaluating the medical images. The authors found that integrating of the three CNN models enhanced the
overall performance of diagnosis system. However, it should be noted that the method employed in the study
needed high computationally demanding. Also the dataset was used included just three forms of skin disorders.
As a result, the architecture that used in this work may not be ideally suited for scenarios that involve multiclass
classification.
Albawi et al. [28] presented a novel method for effectively identifying three specific skin diseases,
namely nevus, melanoma, and atypical. To preprocess skin images, authors utilized an adaptive filtering
technique to eliminate noise. Following that, they employed an adaptive region growing method to precisely
localize and extract the regions of interest (ROI) corresponding to the affected areas. For extracting relevant
features, they employed a hybrid approach that combined two-dimensional discrete wavelet transform with
texture features and geometric. This combination allowed them to capture important information. Lastly, they
implemented CNN on the international skin imaging collaboration (ISIC) dataset. The suggested method
yielded remarkable results, achieving a classification accuracy of 96.768% for the identified diseases.
Ahammed et al. [29] introduced a digital hair removal technique that makes use of morphological
filteration methods, such as black-hat transformation and an inpainting technique. They also applied Gaussian
filteration to address image blurring or noise. The automatic Grabcut segmentation technique was utilized
for accurate lesion segmentation. For extracting relevant patterns associated with images of skin, the
authors employed techniques like statistical features and gray level co-occurrence matrix (GLCM).
Ramachandro et al. [30] focused on classifying skin images, specifically targeting 4 types of skin tumors. They
employed models based on deep learning, particularly transfer learning, utilizing pretrained deep neural
networks such as DenseNet and CNN models. Additionally, they incorporated machine learning methods
inclusing SVM and random forest (RF) classifier in their analysis.
3. METHOD
The objective of the research is to provide an effective model for detection of skin cancer using deep
neural networks. Deep neural networks excel in image-related tasks due to their depth and the efficiency of
convolutional filters. In this study, we employed the AlexNet deep neural network for this purpose. However,
a significant challenge in enhancing the operation of neural networks is the precise tuning of the parameters.
To address this challenge, we utilized the MPA algorithm for the optimal tuning of AlexNet network
parameters. The schematic representation illustrating the presented method is depicted in Figure 1.
3.1. Data preprocessing
During the preprocessing phase, the image size is initially adjusted. As mentioned in the database
description, the image size in the database is 600×450 pixels. However, the input image size for the AlexNet
neural network is 227×227 pixels. Therefore, the image size is first resized to these dimensions. Subsequently,
the data is segmented into test and training sets. In our work, 70% of the dataset is considered for training, and
30% for evaluating the network. The training set is applied to perform the training operation and update weights
of the neural network. The test data is utilized to assess the network efficacy based on accuracy and
generalization capabilities. The steps of the presented model are explained as follows.
3.2. AlexNet deep neural network
In this research, the AlexNet deep neural network is deployed for skin cancer detection. The proposed
architecture of AlexNet includes a set of pooling layers, convolutional layers, fully connected layers, and a
softmax layer. Structure of this network is depicted in Figure 2. This network is described as follows:
− Input layer: Images are provided to the network as input in this layer. Prior to entering the AlexNet
network, preprocessing is applied. This process involves resizing the images to the dimensions required
by the network and normalizing pixel values to a standardized scale (e.g., [0,1]).
− Convolutional layers: AlexNet comprises several convolutional layers that act as filters moving across
images, detecting various features such as edges, enhancers, and different objects. These layers extract
lower-level features like lines and shapes.
− Pooling layers: In the proposed architecture, pooling layers are placed after each convolutional layer.
Pooling layers compress features and use essential information for subsequent stages. These layers help
to reduce the spatial dimensions of the image.

4825
− Fully connected layers: The extracted features enter these layers, subsequent to the pooling and
convolutional layers. The fully connected layers are typically connected to a deeper neural network,
capturing higher-level features and more complex combinations in the images.
− Activation function: In the hidden neural layers, activation functions like rectified linear unit (ReLU) are
used. Activation functions nonlinearly transform the network's performance and extract nonlinear features
from images.
− Output layers: The final result of this stage in the AlexNet network includes predicted labels for each
image. These labels determine the detected class for each image.
Figure 1. The diagram of the presented model
Figure 2. AlexNet deep neural network architecture
3.3. Fine-tuning AlexNet hyperparameters using marine predators algorithm
Utilizing deep CNN such as AlexNet for skin cancer detection can offer an effective solution with
high accuracy. However, the challenge with using neural network models lies in the issues of overfitting or
poor performance, often stemming from suboptimal parameter tuning. In the presented study, the MPA
algorithm is employed to achieve optimal hyperparameter tuning for the AlexNet neural network, aiming to
overcome challenges in the learning process. The hyperparameters of the AlexNet network, which were
optimized based on this algorithm, are outlined in Table 1.

 ISSN: 2252-8938
4826
Table 1. The AlexNet network hyperparameters
Description Parameters
Rate of learning Learning rate
Mini batch Size mini batch
Optimization algorithm (sgdm, rmsprop, adam) Optimizer
Kernel size in Conv. layer Kernel Size
Frequency of validation Validation frequency
3.3.1. Marine predators algorithm
The MPA algorithm is a population-based approach that simulates the steps a marine predator takes
to hunt its prey. In the initial phase, the predator remains stationary. In the next stage, it exhibits Brownian
motion. Then, in the third stage, it operates according to the Levy strategy. This operation is also applicable to
prey, which can be another potential predator. For instance, both sharks and tuna are regarded as marine
predators, with tuna often serving as prey for sharks, it acts as a predator for other fish and invertebrates in the
marine ecosystem. The steps of MPA algorithm are described as follows:
a) Initialization of the initial population
In population-based algorithms, the initial population (initial solution) constitutes a set of potential
solutions to the problem, generated randomly. Here, the initial population consists of a set of initial values for
the hyperparameters of the AlexNet network. Each member of the population, referred to as a search agent,
represents a potential solution with optimal values for the hyperparameters. As the algorithm iterates, the
solution values converge towards optimal values. Therefore, the position of each search agent in the parameter
space of AlexNet hyperparameters needs to be optimized. In the MPA algorithm, the primary population is
uniformly spread in the search space based on (1).
𝑋0 = 𝑋𝑚𝑖𝑛 + 𝑟𝑎𝑛𝑑(𝑋𝑚𝑎𝑥 − 𝑋𝑚𝑖𝑛) (1)
where 𝑋𝑚𝑖𝑛 and 𝑋𝑚𝑎𝑥 are the minimum and maximum limits for the variables, and 𝑟𝑎𝑛𝑑 is a vector of
uniformly distributed random values in the range of [0,1].
b) Prey and elite matrices
The prey matrix represents the current positions associated with the search agents. In this matrix, each
row signifies the current position of the i-th search agent. The matrix of prey is illustrated in (2).
𝑃𝑟𝑒𝑦 = [
𝑋1,1 𝑋1,2 ⋯ 𝑋1,𝑑
𝑋2,1 𝑋2,2 ⋯ 𝑋2,𝑑
⋮ ⋮ ⋯ ⋮
𝑋𝑛,1 𝑋𝑛,1 ⋯ 𝑋𝑛,𝑑
]
𝑛×𝑑
(2)
where 𝑋𝑖, 𝑗 denotes the j-th dimension of the i-th search agent, 𝑛 represents the count of search agents, while d
indicates the number of dimensions.
According to the theory of “survival of the fittest”, elite predators exhibit a higher hunting potential in
nature. Therefore, the best solution of each search agent is selected as an elite predator to create the elite matrix.
𝐸𝑙𝑖𝑡𝑒 =
[
𝑋1,1
𝐼
𝑋1,2
𝐼
⋯ 𝑋1,𝑑
𝐼
𝑋2,1
𝐼
𝑋2,2
𝐼
⋯ 𝑋2,𝑑
𝐼
⋮ ⋮ ⋮ ⋮
⋮ ⋮ ⋮ ⋮
𝑋𝑛,1
𝐼
𝑋𝑛,2
𝐼
⋯ 𝑋𝑛,𝑑
𝐼
]𝑛×𝑑
(3)
where 𝑋𝑖.𝑗
𝐼
⃗⃗⃗⃗⃗⃗ represents the Elite predator vector of the i-th search agent in the j-th dimension. Also, 𝑛 denotes
the count of search agents, and d denotes the dimensions number.
c) Position update
After creating an initial population and calculating the prey and elite matrices, the search agents’
positions need to be updated accordingly. The MPA uses an intelligent approach to balance exploration and
exploitation phases. This method utilizes three different movement strategies to update the locations of the search
agents. The algorithm consists of three distinct phases to update the locations of the search agents, ensuring a
balance between exploration and exploitation operations. These stages are determined in the following:
− The first phase
During the first-third of the search operations (Iteration<1/3*Iteration_max), the agents update their
positions using Brownian motion. The new position is calculated using (4):

4827
𝑊ℎ𝑖𝑙𝑒 𝐼𝑡𝑒𝑟 <
1
3
𝑀𝑎𝑥_𝐼𝑡𝑒𝑟
𝑠𝑡𝑒𝑝𝑠𝑖𝑧𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝑅𝐵
⃗⃗⃗⃗⃗ ⨂(𝐸𝑙𝑖𝑡𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ − 𝑅𝐵
⃗⃗⃗⃗⃗ ⨂ 𝑃𝑟𝑒𝑦𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ) 𝑖 = 1, … , 𝑛 (4)
𝑃𝑟𝑒𝑦𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝑃𝑟𝑒𝑦𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ + 𝑃. 𝑅
⃗ ⨂𝑠𝑡𝑒𝑝𝑠𝑖𝑧𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
Here, the vector 𝑅𝐵
⃗⃗⃗⃗⃗ is composed of random numbers that follow a normal distribution, symbolizing brownian
motion. The constant P is assigned a value of 0.5, and vector 𝑅
⃗ consists of random numbers that are uniformly
distributed within the range of [0,1]. Iter represents the current iteration, and 𝑀𝑎𝑥_𝐼𝑡𝑒𝑟 is the maximum number
of iterations. Brownian motion assists predators in exploring distinct regions around them, leading to effective
exploration in the initial iterations where predators are evenly distributed across the search space, and the
distance between them is relatively high.
− The second phase
In the next one-third of the search operations (1/3*Iteration_max <Iteration<2/3*Iteration_max), half
of the population is allocated for exploration using brownian motion, while the other half uses levy flight for
exploitation. The new positions for half of the population are calculated based on (5).
𝑊ℎ𝑖𝑙𝑒
1
3
𝑀𝑎𝑥_𝐼𝑡𝑒𝑟 < 𝐼𝑡𝑒𝑟 <
2
3
For the first half of the population;
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝑅𝐿
⃗⃗⃗⃗ ⨂(𝑅𝐿
⃗⃗⃗⃗ ⨂𝐸𝑙𝑖𝑡𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ − 𝑃𝑟𝑒𝑦𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ) 𝑖 = 1, … , 𝑛/2 (5)
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝑃𝑟𝑒𝑦𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ + 𝑃. 𝑅
⃗ ⨂𝑠𝑡𝑒𝑝𝑠𝑖𝑧𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
Also, for the other half using levy flight,
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝑅𝐿
⃗⃗⃗⃗ ⨂(𝑅𝐿
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ − 𝑃𝑟𝑒𝑦𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ) 𝑖 = 𝑛/2, … , 𝑛 (6)
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝐸𝑙𝑖𝑡𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ + 𝑃. 𝐶𝐹⨂𝑠𝑡𝑒𝑝𝑠𝑖𝑧𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
Here, the vector 𝑅𝐿
⃗⃗⃗⃗ consists of random numbers that follow a Levy distribution, symbolizing a levy flight. The
variable CF represents an adaptive metric for regulating the step size of the levy flight, calculated as (7).
CF = (1 −
Iter
MaxIter
)
(2
Iter
MaxIter
)
(7)
− The third phase
During the last stage of optimization (Iteration>2/3*Iteration_max), the entire population is allocated
for exploitation.
𝑊ℎ𝑖𝑙𝑒 𝐼𝑡𝑒𝑟 >
2
3
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝑅𝐿
⃗⃗⃗⃗ ⨂(𝑅𝐿
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ − 𝑃𝑟𝑒𝑦𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ) 𝑖 = 1,2, … , 𝑛 (8)
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = 𝐸𝑙𝑖𝑡𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ + 𝑃. 𝐶𝐹⨂𝑠𝑡𝑒𝑝𝑠𝑖𝑧𝑒𝑖
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
In this phase, the predators’ transition from brownian motion to levy flight focus on their best local
search for enhanced exploitation. The adaptive variable CF significantly aids predators in restricting search
ranges in specific neighborhoods for exploitation and preventing wasted efforts due to the long steps of levy
flight in inappropriate domains.
d) Update of elite matrix
Once the positions of the search agents have been adjusted, the fitness value for each agent is
evaluated. If the fitness value for the current solution of the i-th search agent surpasses the previous fitness
value, the current position of that agent will be updated in the elite matrix. This comparison is conducted for
all search agents in each iteration.
e) Fish aggregating devices effect
Research findings indicate that sharks are observed in the vicinity of fish group activities for more
than 80% of their time. For 20% that remains, they perform longer jumps in various dimensions, possibly
seeking environments with different prey distributions. Fish aggregation devices (FADs) are perceived as local
optimum, and their influence is akin to becoming ensnared at these points within the search space. Considering

 ISSN: 2252-8938
4828
the longer jumps during this phase prevents getting trapped in local optimum. Therefore, the FADs effect is
determined as (9).
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ = {
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ + 𝐶𝐹[𝑋𝑚𝑖𝑛
⃗⃗⃗⃗⃗⃗⃗⃗⃗ + 𝑅
⃗ ⨂(𝑋𝑚𝑎𝑥
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ − 𝑋𝑚𝑖𝑛
⃗⃗⃗⃗⃗⃗⃗⃗⃗ )]⨂𝑈
⃗
⃗ 𝑖𝑓 𝑟 ≤ 𝐹𝐴𝐷𝑠
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ + [𝐹𝐴𝐷𝑠(1 − 𝑟) + 𝑟](𝑃𝑟𝑒𝑦𝑟1
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ − 𝑃𝑟𝑒𝑦𝑟2
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗ ) 𝑖𝑓 𝑟 > 𝐹𝐴𝐷𝑠
(9)
The symbol 𝐹𝐴𝐷𝑠 denotes the likelihood of FADs exerting an impact on the operational procedure,
set to 0.2. Also, U
⃗⃗ represents a binary vector with arrays containing zero and one, generated randomly.
Additionally, the variable r represents a random number uniformly distributed in the range of 0 and 1. The
variables 𝑟1 and 𝑟2 correspond to random indexes of the Prey matrix.
f) Memory and elite matrix update
Given the outlined considerations, marine predators derive advantages from a robust memory that
enhances their success in hunting. This ability is emulated through memory preservation in the MPA. Upon
updating the prey and incorporating the FADs effect, the fitness of each solution in the current iteration is
juxtaposed with its equivalent from the preceding iteration. If the current solution's fitness value is higher, it
replaces the preceding solution. The pseudocode outlining the MPA algorithm for optimizing the parameters
of the AlexNet neural network is illustrated in Pseudocode 1.
Pseudocode 1. The MPA algorithm pseudocode
Input: Maximum Iteration Number, initial value of Parameters that must be optimized
(Learning Rate, Mini-Batch Size, Optimizer function, Kernel Size and Validation Frequency),
Permissible range of variables.
Output: Optimized Value of Learning Rate, Mini-Batch Size, Optimizer function, Kernel Size
and Validation Frequency
Initialize search agents’ population
While (maximum iteration number reach)
Calculate the fitness value and generate the Elite matrix.
If Iteration<Max_Iteration/3
Update prey-matrix according to Eq. 4
Else if max_ Iteration /3< Iteration <2*max_ Iteration /3
for first half part of population (i=1,…,n/2)
update prey-matrix according to Eq. 5
for the second part of the population (i=n/2,…,n)
Else if Iteration >2*max_ Iteration /3
End (if)
Updates Elite-matrix according to current fitness and previous fitness that saved in memory
Applying fads effect and update prey-matrix according to Eq. 9
End while
4. THE EXPERIMENTAL RESULTS
In this stage, we provide the simulation outcomes of the suggested model for the diagnosis of skin
cancer. Initially, evaluation criteria are introduced. Subsequently, the findings of the presented method are
analyzed and a comparison with other techniques is performed. It should be noted that simulations are
conducted using Google Colab. Additionally, data are divided into training and test datasets. In this process,
70% data is considered for training and 30% for test and evaluation.
4.1. Dataset
In this study, the HAM-10000 dataset is utilized for skin cancer detection. The HAM10000 database
containing 10,015 dermatoscopy images depicting various skin lesions. The dataset was gathered from patients
in Australia and Austria. The images have dimensions of 600*450 pixels and are centered crops. This database
consists of 7 classes, each representing a specific disease. These classes are reported in Table 2. Some examples
of the database images are illustrated in Figure 3.
Table 2. HAM-10000 database classes
Class No. Disease type
0 Actinic keratoses and intraepithelial carcinoma/bowen disease
1 Basal cell carcinoma
2 Benign lesions of the keratosis
3 Dermatofibroma
4 Melanoma
5 Melanocytic nevi
6 Vascular lesions

4829
Figure 3. Dataset image examples
4.2. Evaluation metric
In our work, precision, accuracy, recall and F-measure are employed to evaluate the effectiveness of
the presented model. The equations for these metrics are presented as follows:
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
(𝑇𝑃+𝑇𝑁)
(𝑇𝑃+𝐹𝑃+𝑇𝑁+𝐹𝑁)
(10)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃
(𝑇𝑃+𝐹𝑃)
(11)
𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃
(𝑇𝑃+𝐹𝑁)
(12)
𝐹1 𝑠𝑐𝑜𝑟𝑒 =
2∗(𝑅𝑒𝑐𝑎𝑙𝑙∗𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
(𝑅𝑒𝑐𝑎𝑙𝑙+𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
(13)
Here, TP represents true positive rate (TPR) of detection, TN denotes true negative rate of detection, FP
signifies false positive rate (FPR) of detection, and FN indicates false negative rate of detection.
4.3. Evaluating training process
Figure 4 illustrates the learning curve in terms of the loss function on the training set for 1000 epochs.
The loss, also referred to as the cost, quantifies the model's error. The objective of the model is to minimize
the loss, which is achieved through techniques such as gradient descent. Hence, as the learning progresses, a
smaller loss indicates improved model performance. To quantify the loss, a loss or cost function is evaluated.
In Figure 4, the expected evolution of the learning process is observed. It is evident that the loss curve
consistently decreases over the course of training. As the learning curve demonstrates, the network loss has
increased from 0.5 per 400 epochs to 0.15 per thousand epochs. This means that the learning process in the
proposed model has occurred correctly.
Figure 4. Learning curve

 ISSN: 2252-8938
4830
4.4. Result evaluation and discussion
As was mentioned, precision, accuracy, recall, and F1 measure metrics are employed to assess the
effectiveness of the presented approach. The numerical findings of the evaluation metrics obtained from
simulations for training and test data are presented in Table 3. As can be seen, the parameters of accuracy,
precision, F1-score and recall for the training data are 99.11, 99.13, 99.11, and 99.11, respectively. Also, the
value of these metrics for the test data is 98.47, 98.51, 98.47, and 98.47, respectively. It is obvious that these
parameters achieved higher values for training data than for test data. Since the network has seen this data in
the training process.
Table 3. Evaluation metric results for train and test data
Metric Results for training data Results for test data
Accuracy 99.11 98.47
Precision 99.13 98.51
Recall 99.11 98.47
F-score 99.11 98.47
Using a range of thresholds, TPR is plotted against the FPR to observe the trade-off between these
two measures is observed. A highly accurate classifier tends to be positioned towards the upper left corner of
the receiver operating characteristic (ROC) curve, exhibiting a high TPR and a low FPR. Conversely, a poorly
performing classifier is typically situated towards the lower right corner of the ROC curve, characterized by a
low TPR and a high FPR. Additionally, a random classifier lies along the diagonal line of the ROC curve,
indicating an equal TPR/FPR ratio. The ROC curve of the proposed model is depicted in Figure 5. As
illustrated, this curve demonstrates a high TPR and a low FPR, positioned near the upper left corner.
Consequently, it can be concluded that the presented method has produced accurate results in the skin cancer
classification. Figure 5 presents the ROC curve.
Figure 5. The ROC curves
Here, we delve into the examination and comparison of the outcomes achieved from the presented
system with other approaches. Table 4 illustrates the comparison of outcomes based on accuracy, recall,
F1 score, and precision metrics. As the tables presents, the presented method is superior to other methods in
terms of all criteria. For example, the suggested approach achieved 98% accuracy, while the highest accuracy
after the presented approach is related to the SVM with a value of 97%. Our study suggests that higher recall
is not associated with poor performance in precision. The proposed method may benefit from recall without
adversely impacting precision. In addition, we found that the recall and precision measures were correlated
with the F-score measure. The proposed method in this study tends to have an extremely high ratio of Fscore
compared to other previous studies. The superiority compared to other techniques is also observed in other
criteria which are given in Table 4. The reason lies in the precise and optimal tuning of the AlexNet parameters
using the MPA.
It is worth mentioning that this research has investigated simultaneously all the criteria of accuracy,
precision, recall, F-measure, ROC curve, and learning curve. While previous studies have not investigated the

4831
simultaneous effect of all these criteria, and previous studies have investigated some of these criteria. In
addition, in this paper a comprehensive study in automatic transfer learning based skin cancer diagnosis using
fine-tuned AlexNet by MPA algorithm explored. However, further and in-depth studies may be needed to
confirm its efficient on more patient. In this study shown that swarm intelligence-based optimization algorithms
such as MPA algorithm are more efficient than traditional methods for fine-tuning deep neural network hyper
parameters. Future studies may investigate more up-to-date meta-heuristic algorithms for tuning hyper
parameters of deep neural networks. Recent observations in the studies of automatic detection of skin cancer
show that deep learning-based methods can provide a good performance in diagnosing this disease with high
accuracy. Our findings provide conclusive evidence that fine-tuning hyper parameters of pre-trained deep
neural network using optimization algorithms based on meta-heuristic methods can lead to better performance
of the skin cancer diagnosis model. The results obtained in this work confirm this issue.
Table 4. Results comparison based on accuracy, precision, recall, and F1 score
Method F-Score Recall Precision Accuracy
KNN [29] 94.71 95.14 95.14 95
DT [29] 95.14 95.57 95.71 95
SVM [29] 97.43 97.57 97.71 97
RF [30] 83 89 91 92
CNN [30] 92 89 91 93
DesNet [30] 83 84 93 95
The proposed method (MPA-AlexNet) 98.47 98.47 98.51 98
5. CONCLUSION
This paper introduces the improved AlexNet neural network with MPA to detect skin cancer. Early
diagnosis of this disease will drastically reduce mortality rates and save lives. In recent years, technologies
based on artificial intelligence such as neural networks have achieved promising results in this field. Our
proposed method involves optimizing the AlexNet network parameters by the MPA algorithm. The results of
the simulations show the superiority of the presented approach based on the evaluation parameters used in the
experiments. AlexNet neural network is powerful in the field of recognizing and classifying patterns from
images. Due to its parallel search capability, the MPA algorithm has a good convergence speed and does not
get trapped in local optima. Therefore, the integration of the AlexNet neural network and the MPA algorithm
in the proposed method has contributed significantly to its outperformance in comparison with other
state-of-the-art methods.
REFERENCES
[1] N. Hameed, A. Shabut, and M. A. Hossain, “A computer-aided diagnosis system for classifying prominent skin lesions using
machine learning,” 2018 10th Computer Science and Electronic Engineering Conference, CEEC 2018 - Proceedings, pp. 186–191,
2019, doi: 10.1109/CEEC.2018.8674183.
[2] U. O. Dorj, K. K. Lee, J. Y. Choi, and M. Lee, “The skin cancer classification using deep convolutional neural network,” Multimedia
Tools and Applications, vol. 77, no. 8, pp. 9909–9924, 2018, doi: 10.1007/s11042-018-5714-1.
[3] H. Bhatt, V. Shah, K. Shah, R. Shah, and M. Shah, “State-of-the-art machine learning techniques for melanoma skin cancer detection
and classification: a comprehensive review,” Intelligent Medicine, vol. 3, no. 3, pp. 180–190, 2023, doi:
10.1016/j.imed.2022.08.004.
[4] D. Wen et al., “Characteristics of publicly available skin cancer image datasets: a systematic review,” The Lancet Digital Health,
vol. 4, no. 1, pp. e64–e74, 2022, doi: 10.1016/S2589-7500(21)00252-1.
[5] K. Hauser et al., “Explainable artificial intelligence in skin cancer recognition: A systematic review,” European Journal of Cancer,
vol. 167, pp. 54–69, 2022, doi: 10.1016/j.ejca.2022.02.025.
[6] M. A. Bagheri, G. A. Montazer, and S. Escalera, “Error correcting output codes for multiclass classification: Application to two
image vision problems,” AISP 2012 - 16th CSI International Symposium on Artificial Intelligence and Signal Processing, pp. 508–
513, 2012, doi: 10.1109/AISP.2012.6313800.
[7] L. Zeng et al., “Advancements in nanoparticle-based treatment approaches for skin cancer therapy,” Molecular Cancer, vol. 22, no.
1, 2023, doi: 10.1186/s12943-022-01708-4.
[8] K. Ali, Z. A. Shaikh, A. A. Khan, and A. A. Laghari, “Multiclass skin cancer classification using EfficientNets – a first step towards
preventing skin cancer,” Neuroscience Informatics, vol. 2, no. 4, 2022, doi: 10.1016/j.neuri.2021.100034.
[9] D. Keerthana, V. Venugopal, M. K. Nath, and M. Mishra, “Hybrid convolutional neural networks with SVM classifier for
classification of skin cancer,” Biomedical Engineering Advances, vol. 5, 2023, doi: 10.1016/j.bea.2022.100069.
[10] M. Dildar et al., “Skin cancer detection: a review using deep learning techniques,” International Journal of Environmental Research
and Public Health, vol. 18, no. 10, 2021, doi: 10.3390/ijerph18105479.
[11] A.V.Ubale and P. L. Paikrao, “Detection and classification of skin diseases using different color phase models,” International
Research Journal of Engineering and Technology (IRJET), vol. 6, no. 7, pp. 3658–3663, 2019.
[12] R. Leon et al., “Non-invasive skin cancer diagnosis using hyperspectral imaging for in-situ clinical support,” Journal of Clinical
Medicine, vol. 9, no. 6, 2020, doi: 10.3390/jcm9061662.
[13] Y. R. Woo, S. H. Cho, J. D. Lee, and H. S. Kim, “The human microbiota and skin cancer,” International Journal of Molecular
Sciences, vol. 23, no. 3, 2022, doi: 10.3390/ijms23031813.

 ISSN: 2252-8938
4832
[14] C. King et al., “Somatic mutations in facial skin from countries of contrasting skin cancer risk,” Nature Genetics, vol. 55, no. 9, pp.
1440–1447, 2023, doi: 10.1038/s41588-023-01468-x.
[15] M. Wada, Z. Y. Ge, S. J. Gilmore, and V. J. Mar, “Use of artificial intelligence in skin cancer diagnosis and management,” Medical
Journal of Australia, vol. 213, no. 6, pp. 256-259, 2020, doi: 10.5694/mja2.50759.
[16] A. A. Malibari et al., “Optimal deep neural network-driven computer aided diagnosis model for skin cancer,” Computers and
Electrical Engineering, vol. 103, 2022, doi: 10.1016/j.compeleceng.2022.108318.
[17] L. R. -Barroso, S. P. -Gutiérrez, C. Yáñez, F. J. B.-Fernández, M. Vilaseca, and S. Royo, “Optical technologies for the improvement
of skin cancer diagnosis: A review,” Sensors, vol. 21, no. 1, pp. 1–31, 2021, doi: 10.3390/s21010252.
[18] B. Shetty, R. Fernandes, A. P. Rodrigues, R. Chengoden, S. Bhattacharya, and K. Lakshmanna, “Skin lesion classification of
dermoscopic images using machine learning and convolutional neural network,” Scientific Reports, vol. 12, no. 1, 2022, doi:
10.1038/s41598-022-22644-9.
[19] D. Adla, G. V. R. Reddy, P. Nayak, and G. Karuna, “Deep learning-based computer aided diagnosis model for skin cancer detection
and classification,” Distributed and Parallel Databases, vol. 40, no. 4, pp. 717–736, 2022, doi: 10.1007/s10619-021-07360-z.
[20] J. P. Jeyakumar, A. Jude, A. G. Priya, and J. Hemanth, “A survey on computer-aided intelligent methods to identify and classify
skin cancer,” Informatics, vol. 9, no. 4, 2022, doi: 10.3390/informatics9040099.
[21] Y. Wu, B. Chen, A. Zeng, D. Pan, R. Wang, and S. Zhao, “Skin cancer classification with deep learning: a systematic review,”
Frontiers in Oncology, vol. 12, 2022, doi: 10.3389/fonc.2022.893972.
[22] V. Anand, S. Gupta, A. Altameem, S. R. Nayak, R. C. Poonia, and A. K. J. Saudagar, “An enhanced transfer learning-based
classification for diagnosis of skin cancer,” Diagnostics, vol. 12, no. 7, 2022, doi: 10.3390/diagnostics12071628.
[23] A. G. Diab, N. Fayez, and M. M. El-Seddek, “Accurate skin cancer diagnosis based on convolutional neural networks,” Indonesian
Journal of Electrical Engineering and Computer Science, vol. 25, no. 3, pp. 1429–1441, 2022, doi: 10.11591/ijeecs.v25.i3.pp1429-1441.
[24] S. S. Chaturvedi, K. Gupta, and P. S. Prasad, “Skin lesion analyser: an efficient seven-way multi-class skin cancer classification using
mobilenet,” Advances in Intelligent Systems and Computing, vol. 1141, pp. 165–176, 2021, doi: 10.1007/978-981-15-3383-9_15.
[25] K. M. Hosny, M. A. Kassem, and M. M. Foaud, “Classification of skin lesions using transfer learning and augmentation with Alex-
net,” PLoS ONE, vol. 14, no. 5, 2019, doi: 10.1371/journal.pone.0217293.
[26] T. Shanthi, R. S. Sabeenian, and R. Anand, “Automatic diagnosis of skin diseases using convolution neural network,”
Microprocessors and Microsystems, vol. 76, 2020, doi: 10.1016/j.micpro.2020.103074.
[27] R. Bhavani, V. Prakash, R. V. Kumaresh, and R. Sundra Srinivasan, “Vision-based skin disease identification using deep learning,”
International Journal of Engineering and Advanced Technology, vol. 8, no. 6, pp. 3784–3788, 2019, doi:
10.35940/ijeat.F9391.088619.
[28] S. Albawi, Y. A. Abbas, and Y. Almadany. “Robust skin diseases detection and classification using deep neural networks,”
International Journal of Engineering & Technology, vol. 7, pp. 6473–6480, 2018.
[29] M. Ahammed, M. Al Mamun, and M. S. Uddin, “A machine learning approach for skin disease detection and classification using
image segmentation,” Healthcare Analytics, vol. 2, 2022, doi: 10.1016/j.health.2022.100122.
[30] M. Ramachandro, T. Daniya, and B. Saritha, “Skin cancer detection using machine learning algorithms,” 3rd IEEE International
Virtual Conference on Innovations in Power and Advanced Computing Technologies, i-PACT 2021, 2021, doi: 10.1109/i-
PACT52855.2021.9696874.
BIOGRAPHIES OF AUTHORS
Maha Ibrahim Khaleel is an assistant lecturer in the Department of Computer
Technology Engineering at Al-Safwa University College in Karbala from 2020 to now. She is
also now a doctoral student at Qom University in Qom. Her research interests include soft
computing, deep learning, smart systems, and the internet of things (IOT). She can be contacted
at email: m.ibrahim@stu.qom.ac.ir or maha.ibrahim@alsafwa.edu.iq.
Amir Lakizadeh is an assistant professor in the Department of Computer
Engineering and Information Technology of University of Qom. He received his doctorate in
computer engineering and in the last 20 years, he has researched the application of machine
learning methods and especially deep learning in bioinformatics, computational biology, disease
diagnosis, drug repurposing, and drug design. He can be contacted at email: lakizadeh@qom.ac.ir.

Transfer-learning based skin cancer diagnosis using fine-tuned AlexNet by marine predators algorithm

More Related Content

Similar to Transfer-learning based skin cancer diagnosis using fine-tuned AlexNet by marine predators algorithm (20)

More from IAESIJAI (20)

Recently uploaded (20)

Transfer-learning based skin cancer diagnosis using fine-tuned AlexNet by marine predators algorithm