Adaptive Fine-Tuned AdaBoost and Improved Firefly Algorithm for Skin Cancer Detection

Melanoma is among the most frequently pinpointed and extremely dangerous carcinomas currently being investigated. The irregular development of skin cells is what is known as skin cancer. If left untreated, the cancerous cells in this area of the skin might spread to the surrounding living tissue. Many morphological characteristics, including color systems, dots, Streaks, blue-white regions, and blotches, can be identified. Squamous cell carcinoma (SCC), basal cell carcinoma (BCC) & malignant melanoma (MM) are indeed the 3 most prevalent kinds of tumors. The worst kind of skin cancer is malignant melanoma (MM). It typically begins as a mole and then develops into a large dark area on the skin. It is critical to diagnose and treat in the worst cases [1].

After the excision of the lesion, there is a possibility of its recurrence. To precisely recognize skin conditions for cancer diagnosis, a mere visual examination of the skin is inadequate. The inclusion of a scheme in the diagnostic test for cancer detection is considered crucial. This scheme is responsible for various tasks such as picture pre-processing to improve contrast by removing artifacts and noise, segmentation to separate the lesion from healthy tissue through boundary detection, removal of features to enhance the speed and accuracy of supervised methods, and categorization by labeling image data into predefined classes. The uneven margins, structure, thickness, and color change of the lesion make it hard to identify it from healthy skin in image data. The appearance of artifacts on the lesions would make segmentation that much more difficult. Artifacts, including bubbles, hair, ink marks, spots, and pores, produce abnormalities in dermoscopic pictures, leading to improper segmentation during skin lesion assessment. As a result, preparing dermoscopy pictures to eliminate image capture distortion, texture, color variations, and unwanted structures like hairs, air bubbles, and blurs surrounding the lesion is critical in determining the quality of lesion diagnosis. The non-invasive nature of image recognition has made it a popular method for identifying melanoma skin cancer, and as a result, fast and suitable therapy can be given to the patient.

As a result, identifying new or altering lesions that could be malignant is critical. Skin tumors that are discovered and removed early on are virtually always curable. Pre-processing is necessary for skin cancer testing. Most of those methods are well-liked [2]. The process of fragmenting the desired object region is defined as segmentation. The process of extracting features plays a crucial role in acquiring a discerning representation of skin lesions. Identifying valuable attributes is a challenging endeavor, and extensive research has been conducted in this field, enabling the recognition of a diverse array of features that define skin lesions [3]. The selection of features is the difficulty of deciding which features are essential and adequate to represent a notion. Feature classification is a difficult task that necessitates the careful involvement of multiple variables. Selecting training data, processing images, extracting features, choosing appropriate classification techniques, and then assessing performance are all significant processes in image classification [4].

The following are the contributions of this work:

• The skin images are preprocessed using MF and CIE techniques to homogenize all true skin images and improve their quality, respectively.
• During the segmentation process, the data is broken up into smaller groups called "subclasses." IV3-CA sparse representations with discriminative light codes to characterize features are provided for skin lesions.
• To extract the proportions of skin lesions in the feature extraction stage, IRV2 is implemented.
• To identify dependable feature subsets, the proposed IFFA which fits the characteristic of the fireflies' approach is employed.
• In continuation to achieve the optimal classification of cutaneous tumors, the recommended AFTAA technique is implemented.

The additional details of this research were organized into the following parts: Part 2 explains the literature survey and problem statement. Part 3 explains the presented approach. Part 4 displays the findings and performance investigation. Part 5 summarizes the overall research.

Table 1. Literature summary of various methods

Ref	Dataset	Model	Highlights	Limitations	Acc(%)
[38]	ISIC Archive	VGG	Investigated four distinct data augmentation techniques and a multi-layer augmentation technique for the classification of melanoma.	The data augmentation techniques examined in this research have certain limitations and haven't been tested on a sizable number of datasets.	82.9
[39]	Private dataset	Resnet34, ResNet50 ResNet101 ResNet152	Methods for enhancing dermoscopy categorization using deep learning and producing datasets have been suggested.	Information from further modalities is not taken into account, such as the patient's medical background or details on additional symptoms.	85.0
[40]	IAD dataset	VGG19	The VGG-19 architecture was initially used to assess the thickness of melanoma.	The clinical significance of accurately predicting melanoma thickness has led to a shift away from using pre-training approaches for comparison.	87.2
[41]	HAM10k ISIC2019	ResNeXt, SeResNeXt, DenseNet Xception, and ResNet	To determine the best ensemble learning algorithms for skin cancer categorization, a grid search strategy was used.	The current quantity of training data remains inadequate, and most of the networks utilized in ensemble learning are derived from identical models.	88.0
[42]	PAD-UFES	App linked to CNN	To address the issue of data imbalance, two evolutionary algorithms were developed, and weighted loss function and oversampling were also used.	To boost the performance even more, a bigger dataset was required.	92.0
[43]	ISIC-2018	GAN	CGANs were used to extract important data from all layers and create skin lesion images with a variety of textures and shapes while maintaining the training data's stability.	The quantity of data utilized for training purposes was relatively restricted.	94.1
[44]	ISIC-2018	GAN	Skin lesion-specific customizations were made to the proposed GAN framework. Additionally, by tweaking the GAN network's progressive growth structure of the generator and discriminator, it may generate higher resolution and more diversified skin disease images.	The synthetic dataset generated by the GAN did not exhibit sufficient complexity and diversity in comparison to the original dataset.	95.2
[45]	ISBI-2017, PH2	Denseunet Network	An adversarial training strategy combined with an attention module is proposed to improve the model's robustness in skin disease classification and segmentation.	The model continues to exhibit under and over-segmentation as a result of the sparse training data and hazy boundaries of skin disease images.	96.8

1.png

Figure 1. Displays the proposed framework

2.jpg

Figure 2. ISIC skin dataset

Table 2. Dataset ratio

3.3 Pre-processing with MF and CIE

The preprocessing stage allows for the improvement of an image's clarity and precision. The collected dataset's noise can be removed using the MF, and the brightness and contrast have been improved using CIE.

3.3.1 MF

The MF is a highly effective non-linear filtering technique renowned for its ability to preserve image information. The performance of the MF is influenced by the size of the filter window. While a smaller window retains the features, it also results in a decrease in noise reduction. Larger windows possess a considerable capacity for noise suppression, albeit at the expense of maintaining image quality. In this, the neighboring noisy pixels' median value is used to replace the targeted noisy pixels as shown in Eq. (1) and in Figure 3.

$I[i, j]=$ median $(\mathrm{p}[\mathrm{x}, \mathrm{y}],(\mathrm{x}, \mathrm{y}) \in w)$                    (1)

where, $w$ indicated the weightage of the surrounding pixels and $[i, j]$ is the point at the median of an image. The primary function of the filter is to organize the values of a pixel within the image slice in increasing order and then replace the selected point with the value of the middle point. (If there is an equal number of pixels to its neighborhood, the mean of the center two pixels is utilized.) The pseudocode is as follows:

    for (p1=0; p1<n; p1++)

        for (p2=0; p2<n; p2++)

           if(I[p1] < I[p2])

            {

             w=I[p1]

              I[p1] = I[p2]

              I[p2] = w

             }

3.png

Figure 3. Median filter

3.3.2 CIE

Contour lines were worn to represent the shape of a region on a two-dimensional map. The borders of the melanoma are captured using CIE. The contour of the binary image has the lesion area deleted. The digital images of the affected parts and the original images are then combined to create the exact images and skin lesions. It could move in both spatial and temporal directions. Since it helps to increase contrast, this method is especially helpful in the field of medical imaging, especially when the contrast between the ROI and its surroundings is similar. The disparity of the image is described as a parameter using the contrast augmentation index (CAI) formula in Eq. (2) and Eq. (3).

$\mathrm{CAI}=\frac{c_{\text {processed }}}{c_{\text {actual }}}$                   (2)

where,

C_processed = Considered image contrast value

C_actual = Original image contrast value

$\mathrm{IAC}=\frac{m-s}{m+s}$                    (3)

where,

IAC=ImageAreaContrast

m=Foreground image gray-level value

s=Background image gray-level value

3.4 Image segmentation using Inception v3 Clustering Algorithm (IV3-CA)

Image segmentation is the technique of separating an image into main parts. The primary purpose of segmentation was to separate sections of strong correlation and regions of interest (ROI). An important crucial phase of skin lesion dermoscopy images after preprocessing is ROI extraction. For skin segmentation to be successfully categorized, good extracted features and ROI removal are necessary.

Certain clustering methods, such as k-means, exhibit subpar performance in high-dimensional spaces, such as those found in real-world images, where data points are located on complex submanifold areas and different clusters can’t distinguish with the colors and image composition. Initially, the images are processed using a trained iteration of the Inception v3 neural network. It is known for using inception modules designed to learn a combination of local and global features from input data. Now, these extracted features are used for clustering. The Inception architecture hypothesized clustering the similar sparse nodes into a dense structure. It generates chances supported by the evidence shown in Eq.4 retrieved from the pictures. This implementation of dimensionality reduction techniques facilitates the clustering process by simplifies the cluster task. The network produces results that possess significant semantic value. The neural activations within this layer are utilized for final clustering.

By utilizing various filter sizes on the same stage, the Inception v1 network [30] (the first edition of Inception) can be used to handle the problem of change and positioning of important components in an image that applies a technique of "widening" rather than "deepening”. Many of the advantages of v1 are present in IV3-CA, along with improvements from v2. Label softening factorized 7×7 convolutions to minimize computation effort and limit overfitting in its supplementary classifications. We have used an Image Net to train the IV3-CA pre-trained framework. The system fine-tuning technique is largely identical to that of the VGG-16 however there are some minor changes. It employs a functional API instead of VGG-16, which allows for further flexibility when creating complicated techniques with various inputs & outputs. After the network was established, the addition of a convolutional layer coupled with an activation function Rectified Linear Unit (ReLU), a global spatial average pooling layer, a fully connected layer (also with RELU activation), and finally a logistic layer (with sigmoid activation), much like a supermodel. The supermodel, which was trained uses an input image size of (299,299) which is cooperative with the architecture. The learning rate(lr) was set to 0.0001 using the RMSprop optimizer. The dataset was split into train, validation, and test sets with a ratio of 70:15:15 percentage of images respectively. Dropout was set to 0.5. We trained the model for 30 epochs with a batch size of 32. We were capable of fine-tuning convolutional layers of IV3-CA using the training model. For the adjustments to be impacted, the system was modified utilizing SGD with just a low learning speed and momentum. Figure 4 displays the IV3-CA design.

SoftMax analysis was used to retrain the final layer in Inception, where we generate probabilities based on the evidence extracted shown in Eq. (4) retrieved from the pictures. The evidence is calculated based on a sum of weights detected by the intensity of pixels, with added bias.

$Evidence _r=\Sigma_s W_{r s} M_s+b_r$                     (4)

where, $M=$ input, $r=$ class, $W=$ weight, $b=$ bias, and $s=$ index. After that, we estimate the chances by running the evidence through the softmax function shown in Eq. (5).

$Outcome (y)= softmax ( Evidence )$                       (5)

4.png

Figure 4. Inception v3 clustering architecture

5.png

Figure 5. Inception ResNet v2 (IRV2) architecture

3.5 Feature extraction using inception ResNet v2 (IRV2)

The Inception ResNet v2 architecture network was trained on our dataset, while all trainable parameters were fine-tuned across all layers.

This was achieved by fine-tuning the model across all layers and replacing the top layers with a global average pooling layer, a fully connected layer, and a softmax layer. These modifications enabled the classification of the data into two diagnostic categories. The size of input images was all resized to (299,299) to be compatible with this model. The learning rate was set to 0.0001 and ReLU was used for the optimizer. The Soft Attention (SA) block is introduced for the Inception ResNet C block of the system in IRV2 in which the picture size is (8×8). This mechanism highlights the important input parts of an image. It increases the value of crucial characteristics and reduces the impact of disruptive features [41].

Throughout this scenario, the soft attention phase is defined by a (2×2) Maxpooling layer, which will then be combined with the inception building's filter concatenate layer. After the concatenate layer, there is a ReLU activation block. The activation block is followed by the (0.5) dropout layer to regulate the outcome of the attention layer that was depicted in Figure 5. The batch size is set to 16. The convolution procedure of the input data is used to extract the features. The calculation utilized is as follows in Eq. (6):

$V_s^l=\mathrm{f}\left(\Sigma_{r \in m_s}\left(K_{r s}^l \times V_r^{l-1}+b_s^n\right)\right)$                    (6)

where, k=kernel number, V= vectors, b= bias, m=collection of every layer characteristic map.

Below is the summary model of IRV2 with SA shown in Figure 6.

6.png

Figure 6. Summary model of IRV2

3.6 Feature selection using Meta-heuristic approach (FSMH)

The benefits of utilizing a feature selection strategy are shown in Figure 1 and include the increasing ability of a classifier to forecast, obtaining a quick and very efficient gainer, and offering an elementary classification approach. The proposed IFFA modeling with ambient boosting (Ada-Boosting) outperforms previous state-of-the-art Firefly versions in solving several challenging uni-modal, multi-modal minimization, and ensemble minimization challenges. Additionally, the quality of the generated ensemble classifier is superior to the earlier, thorough ensemble models. As a result, using the feature selection method in classification has the major benefit of producing models that are easier to comprehend. In this work, a subset for the prediction of skin cancer was created utilizing the feature selection using the meta-heuristic (FSMH) method. As a result, a technique utilized in feature selection is called an IFFA.

3.6.1 IFFA

One of the newest optimization methods, IFFA, was created by studying how fireflies behave. It's a meta-heuristic technique with naturalistic origins. The algorithm is based on three key traits of fireflies. The gender of fireflies, which is known to be unisex, is one of these traits. Each firefly can therefore be drawn to any other firefly by them. Second, there is an inverse relationship between the distance of the fireflies and their attractiveness. The attraction of two fireflies will be strongest if there is little space between them. The objective function judges a firefly's brightness, which brings us to our final point. According to each problem, Eq. (7) can vary, as illustrated in IFFA Algorithm 1.

$B(d)=\beta_0 e^{-\gamma^{d^2}}$                     (7)

Eq. (7) can be used to calculate the attractiveness of fireflies. Gamma is the firefly's d-dimensional attractiveness. When the distance is 0, the attraction is represented by the fixed light consumption coefficient, beta, which is commonly assumed to be 1. Eq. (8) expresses attractiveness, with 0 representing the starting point at (d = 0), and Eq. (9) representing the updating value

$B=\beta o \exp \left(-\gamma \cdot r^m, j\right), m>=1$                      (8)

$X_i=X_i+B(d) *\left(X_p-X_i\right)+\alpha$                        (9)

Using the attraction formula as indicated in Eq. (8), it is calculated where the less-shiny fireflies will go to join the shiner ones. $\alpha$ and $rand$ are random values that are evenly produced numbers in the range [0, 1] which are used in the equation. The i^th and p^th fireflies in the samples are $x_i$ and $x_p$. The Euclidean Distance (ED) formula, which is depicted in Eq. (10), can be used to determine the distance between two fireflies.

Euclidean distance formula

$d_{i j}=\left\|x_i-x_j\right\| x=\mid \sqrt{\sum_p^d\left(x_{i, p}-x_{j, p}\right)^2}$                  (10)

The proposed IFFA targets two important aspects:

• Reduce the incidents to shorten the optimization period.

• A strict focus on local optima to focus on feasible solutions as possible

Algorithm 1: IFFA [42]

Step 1: Start

Step 2: The Objective functional prototype f(x),

where x= (x1, .........., xd) T

Step 3: Start the pop samples with n flies.

Step 4: Calculate the intensity of light; that is related to f (x)

Step 5: While (s<=MaxGen)

Step 6: Determine the absorptivity coefficient

for C1 = 1: C, for every cluster 1, ................, C

for C2 = 1: R, Execute IFFA R numb of repetitions (with n <= s)

for i=1: n (n Fireflies)

for j=1: n (n Fireflies)

Step7: If (I_j > I_i)

Move ffly i in the direction of j.

 end if

Step 8: Alter the attraction distance d via exp ($-\gamma^{d^2}$)

Step 9: Evaluate recent data and modify light intensity

       end for j

end for i

Step 10: Evaluate recent data and update light intensity

           ( I₁ , ..., m′)

every ffly ( X₁ , .................................., m′)

end for loop C2

Step 11: The flies congregate and depart with the optimal

   X local

end for loop C1

Step 12: By ranking every firefly, find the current global best.

end while

3.7 Ensemble skin cancer classification

The ensemble classifier is generally the average or mean combination of predictions. The ensemble is a form of supervised learning system that can be constructed and shown to predict outcomes. Bagging and Boosting are two popular ensemble algorithms.

3.7.1 Bagging

An ensemble classifier—which frequently consists of a weighted or averaged combination of predictors—performs better than a single classifier. It is possible to develop an ensemble, a type of supervised learning system, and discover that it is capable of prognostication. There are many ensemble algorithms. The most well-known ensemble algorithms are Bagging and Boosting.

3.7.2 Boosting

The performance of a classifier will be assessed by how well it can boost the corresponding F-measure. The AFTAA technique is an improved technique and a leading model of enhancement. The basic idea behind it is to train a lot of weak classifiers providing general abilities on the train set, then merge the weak classifiers using a combination approach to create an active(strong) classifier with better identification capabilities. The most well-known boosting algorithm [20] is called "AdaBoost ," and it uses decision trees as its weak classifiers. More specifically, the (AdaBoost ) functions work as below: Samples for a train set are formulated in Eq. (11).

$T_d=\left\{\left(i_1, t_1\right),\left(i_2, t_2\right), \ldots\left(i_n, t_n\right)\right\}$                  (11)

where, $T_d$  = training data, i = input data, and t = data type.

Step 1: Initialize the weights for every train dataset with N training data points. Weights are given by the AFTA algorithm shown in Eq. (12). Initially every component is given a constant(uniform) weight.

$d(I)=\left(w_{11}, w_{12}, w_{13}, \ldots, w_{1 n}\right)$                     (12)

Step 2: Train the information with the prior or existing weights. The weak classifier being trained should possess an accuracy exceeding 0.5, indicating superior performance. The weak classifier A_r(a) shown in Eq. (13) is counted.

$A_r(\mathrm{a}): \mathrm{a} \rightarrow(-1(0),+1(1))$                    (13)

where, r = 1, 2, . . .n represents the r-th run(cycle).

Step 3: Consider the error level e_r and importance of each weak model A_r(a) shown in Eq. (14):

$\begin{aligned} & e_r=P\left(A_r\left(i_j\right) \neq t_j\right) \\ & \sum_{j=1}^n w_{r j} I\left(A_r\left(i_j\right) \neq t_j\right)\end{aligned}$                    (14)

where, $e_r$= False categorized samples by A_r(a). To transform the problem into a regression task, set true classes to 1 whereas false classes to -1 as shown below in Eq. (15).

$I\left(A_r\left(i_j\right) \neq t_j\right)= \begin{cases}1 & A_r\left(i_j=t_j\right) \\ 0 & A_r\left(i_j \neq t_j\right)\end{cases}$                   (15)

A weight is updated for each observation based on the $e_r$. Eq. (16) shows the recalculating of the weights with a weak classifier.

$f_r=\frac{1}{2} \log \left(\frac{1-e_r}{e_r}\right)$                  (16)

The importance of A_r(i) in the last active(strong) classifier is given as f_r. Its $e_r$ is indirectly proportional to f_r.

Step 4: Normalize the weight dispersion d(r + 1) of the training sample according to f_r. Repeat this process until the number of iterations exceeds the predetermined value “n” as below Eq. (17)

$d(r+1)=\left(w_{r+1,1}, w_{r+1,2}, \ldots w_{r+1, n}\right)$                   (17)

where, $w_{r+1,1}=w_{r+1} / N_r \exp \left(-f_r t_j A_r\left(i_j\right)\right)$, j=1,2,..,n, and N_r is a normalization factor. This factor is denoted in Eq. (18).

$N_r=\sum_{j=1}^n w_{r j} \exp \left(-f_r t_j A_r\left(i_j\right)\right)$                  (18)

Step 5: Merge the weak classifier with a fusion technique to produce the final active(strong) classifier A_a as shown in Eq. (19).

$\mathrm{A}_{\mathrm{a}}=\operatorname{sign}\left(\mathrm{f}_{\mathrm{a}}\right)=\operatorname{sign}\left(\sum_{j=1}^n f_r A_r(a)\right)+\mathrm{T}(\mathrm{f})$                    (19)

where, f_a = sequential function of the weak classifiers across individuals. The factor of the r-th cycle is denoted by f_r. Each weak classifier cycle is indicated as A_r(a). T(f) is a parameter for fine-tuning and is used to note the classified outcome with a higher degree of accuracy.

Early detection of cancer progression is essential for classifying and treating it, and it may even save a life. As a result, from a wide variety of medical data, expert computer algorithms may treat cancer without the intervention of humans. To identify skin cancer using computer vision, the AFTAA approach is proposed in this research. Nearly 2000 photos from the ISIC-2017 were gathered and analyzed using the proposed approach. Additionally, the dataset was correctly classified as benign and malignant using the suggested technique.

It has been demonstrated that the methods employed in classification may deal with the common problem of noisy data (a) by learning descriptive archetypal examples of the two classes (malignant and benign) (b) Segmentation, key point detection, and other appropriate localization techniques are used to extract features, and (c)enhanced procedures are used to get the desired results (selection, classification). Finally, the performance of the suggested technique was examined and compared to cutting-edge procedures. It has been determined that the IFFA with AFTAA has outperformed the other methods.

In addition to the current approaches, the suggested method showed higher accuracy (97.14%), specificity (94.89%) and sensitivity (93.4%). This method, which includes improved pre-processing, may be applied in the future to several datasets for lesion identification and medical diagnostics. Finally, we provide a publicly available dataset. Future research shall focus on expanding to the 3D imaging domain, which could render it beneficial in other medical disciplines as well.

The authors declare that they have no conflicts of interest.