MSCDNet-based multi-class classification of skin cancer using dermoscopy images

Problem statement

Most of the prior deep learning models produce effective results. However, it has some limitations for effective classification, such as it requires numerous steps for analyzing the data, the retrieved handcrafted features cannot guarantee higher classification performance, and poor multi-class classification problems are attained due to an imbalance in data distribution. We propose a new automatic deep learning-based MSCDNet model for accurate classification and segmentation of multiple skin cancer classifications for solving this problem.

Methodology

For multiple skin cancer classification, a new MSCDNet deep learning model is proposed in this article. In the training of the proposed model, ISIC 2019 dataset is utilized. First, performs the pre-processing operation in the dataset images to match the input dimension of the proposed model. Input image sizes are fixed to 224 × 224 resolution for this reason. Additionally, the normalization process is helpful in the images for preventing the overfitting of the model. For detecting skin cancer disease, the proposed model has four primary stages: (1) Pre-processing, (2) selecting the features, (3) classification, (4) segmentation, and (5) performance evaluation. The pre-processing is performed by using a Laplacian-based algorithm and a median filter. Using a Laplacian-based algorithm, the hair is removed from dermoscopy images. Then using a median filter, the unwanted noise from the dermoscopic images is removed. Once the images were pre-processed, we fed them into the feature selection phase to decrease complexity time. With the help of the BDA algorithm, the selection of valuable features is done. The next phase is classification. The obtained features are sent into the deep learning classifier SqueezeNet for multiple skin cancer classifications. This classification network classifies the dermoscopy images into basal cell carcinoma (BCC), benign keratosis (BKL), melanoma (MEL), and melanocytic nevi (MNV). After classification, DenseUNet is employed to segment the defective area. The overall structure of the proposed MSCDNet model is shown in Fig. 1.

Dataset description

The ISIC 2019 was the most often used dataset in this area of research. At https://challenge.isic-archive.com/data#2016, the dataset is obtained. Due to the uneven distribution of images among the eight groups in the ISIC 2019 dataset, it is among the most challenging datasets to classify. With multiple skin cancer classes, the dataset includes 25,331 dermoscopic images. A total of 2,624 benign keratosis images, 3,323 basal cell carcinoma images, 12,875 melanocytic nevi images, and 4,522 Melanoma images are present in the dataset. The rest of the other class images in the dataset are not considered. For training, testing, and validation sets, the dataset is divided into a ratio of 70%:20%:10. A description of the dataset is shown in Table 1. The sample images from the ISIC 2019 datasets are displayed in Fig. 2.

Data augmentation

Deep learning models require a lot of data to function correctly and generalize. Data augmentation techniques include flipping, vertical and horizontal shifting, zooming, random shearing, aspect ratio, pixel jittering, adjusting contrast, modifying brightness, random cropping, flipping, and rotation. Instead of acquiring new data, intentionally increasing the quantity of data is known as “data augmentation.” which is already accessible by adding copies of training data that have been significantly altered. Oversampling or data warping is used to purposefully increase the training dataset’s size or prevent the model from over-fitting from the very start. By making slightly modified copies of previously existing data or by generating synthetic data from previously existing data, it is utilized to increase the diversity of the data.

Rich and poor sets were subtly separated from the ISIC 2019 dataset to improve it relatively. A total of 23,344 examples from the “BCC,” “BKL,” “MEL,” and “NV” cancer classes make up the rich set. The poor set, on the other hand, has 1987 instances of the “AK,” “DF,” “SCC,” and “VASC” classes. As a result, the rich set takes up 92.16% of the dataset, whereas the poor set only accounts for 7.84% of the dataset. To balance the training dataset’s imbalance, augmentation parameters should be applied to the poor but not the rich set classes. To save resources and to label time at the risk of lengthening training time, the online augmentation process is performed for a poor set augmentation process.

Pre-processing

First, the input images are pre-processed using the Laplacian-based algorithm and the median filter. Before using the dermoscopic input images as input for segmentation, hair is first removed since it can significantly or partially conceal some types of melanoma. After converting the image to grayscale, the Laplace operator-based sharpening filter is employed. Then the filtered image is subtracted from the original image. By utilizing a median filter, unwanted image noise is removed. When it comes to keeping an image’s edges, the median filter performs admirably. When the hair has been removed, noise is removed from the images using a median filter.

In the dataset, the dermoscopy images are characterized by 450 × 600-pixel resolution. We reduced the image resolution to 224 × 224 pixels to standardize them with the model input. Additionally, the proposed approach’s proper training was ensured by using the data normalization technique. We then prepared the datasets for MSCDNet input for training. During the pre-processing phase, Fig. 3 shows how well the noise and hair removal procedures worked.

Noise removal process using the median filter

Due to the nonlinear nature of the median filter, the mathematical analysis of the image with random noise is rather challenging. The average filtering’s noise variance is ${\displaystyle {\sigma }_0^2=\frac{1}{n}{\sigma }_i^2}$

where input noise power is represented as ${\sigma }_i^2$ the size of the median filtering mask size is represented as n, the noise density function is depicted as $f\left(\overline{n}\right)$. The effects of median filtering depend on two factors, noise distribution, and mask size. Random noise reduction’s median filtering performance outperforms the average filtering performance. It is quite efficient to use the median filter.

Feature selection

The feature selection process before the use of classification algorithms is important since it improves the classifier’s overall performance, decreases the probability of algorithm failure, and reduces algorithmic computing complexity. Effective classification results are typically achieved by combining multiple features. To improve a classification model’s capacity for discrimination, feature selection is required. For the feature selection process, the redundant or irrelevant features are reduced by discovering the most significant subset of the initial feature set.

This work aims to solve the problem of feature selection by using an algorithm to select features. To accomplish this, the BDA is the primary method used. In 2016, Mirjalili presented the BDA; DA’s discrete variation is called BDA. The relations of dragonflies in identifying the food source (the best solution) and avoiding the enemy (the worst solution) serve as a model for the exploitative and exploratory DA processes. Updating a position in BDA uses five significant behaviors: distraction, attraction, cohesion, alignment, and separation.

The following are descriptions of each of these behaviors.

The static collision between the present individual and a nearby individual is prevented by the Separation process. Separation can be described mathematically as follows: (1)${\displaystyle S{N}_i=-\sum _{j=1}^{N}Q-Q_j}$ where the number of neighbor individuals is represented by N, a dragonfly’s position is indicated by Q, and the neighbor individual’s position is described as Q_j

A swarm or sub-swarm of individuals can match their velocity because of the alignment process. The alignment is determined as follows: (2)${\displaystyle AG{N}_i=\frac{\sum _{j=1}^N{V}_j}{N}.}$

The velocity of neighbor individuals is represented by V_j, and the number of neighbor individuals is represented by N.

Cohesion is the current individual’s divergence from the mass of nearby individuals toward its center. The cohesion is defined as follows: (3)${\displaystyle CH{N}_i=\frac{\sum _{j=1}^N{Q}_j}{N}-Q}$ where j-th dragonfly’s position is indicated by Q_j, and the overall neighborhood dragonfly number is described as N.

Individuals in swarms in nature draw toward food sources and divert predators in addition to separating, aligning, and forming a cohesive group.

According to attraction, the individual must be attracted to possible food sources. The attraction is denoted mathematically as: (4)${\displaystyle F_i=Qf-Q}$ where the position of a food source is indicated by Qf

The individual should be kept away from a potential predator is implied as a distraction. Calculations for the distraction are as follows, (5)${\displaystyle D{T}_i=Qe+Q}$ where the position of the enemy is depicted by Qe

These five behaviors in DA control the movement of dragonflies. The following step vector is calculated for updating the position of each dragonfly: (6)${\displaystyle \Delta Q_i\left(t+1\right)=\left(sS{N}_i+aAG{N}_i+cCH{N}_i+f{F}_i+eD{T}_i\right)+w\Delta Q_i\left(t+1\right)}$ where the separation weight, alignment weight, cohesion weight, food weight, predator weight, and inertia weight are represented as s, a, c, f, e, w, and the current iteration is represented by t.

The following equation is used to update the dragonfly positions in the original DA: (7)${\displaystyle Q_i\left(t+1\right)=Q_i\left(t\right)+\Delta Q\left(t+1\right).}$

This method can handle significant challenges because of such movement and navigation. The following equations are used to update the position vectors in BDA, unlike DA: (8)${\displaystyle Q_i^d\left(t+1\right)=\left\{\begin{array}{c}1-Q_i^d\left(t\right)rand<TF\left(\Delta X_i^d\left(t+1\right)\right)\hfill \\ Q_i^d\left(t\right)rand\ge TF\left(\Delta X_i^d\left(t+1\right)\right)\hfill \end{array}\right.}$ (9)${\displaystyle TF\left(\Delta Q\right)=\left|\frac{\Delta Q}{\sqrt{\Delta Q^2+1}}\right|}$ where $Q_i^d$ is the i-th dragonfly’s position in the d-th dimension, the step vector is represented as ΔQ, the transfer function is TF (.), and the randomly generated number is described as rand in the interval of 0 and 1. The current iteration is represented as t, as shown in Eq. (9).

During the optimizations, several types of local and global searches are provided by BDA with cohesion, alignment, and separation. The other elements that enable the dragonflies to take advantage of the best solutions and avoid the wrong solutions are attraction and distraction. The BDA algorithm is superior due to these five swarming behaviors.

The changing probabilities for the position of dragonflies are determined by the V-shape transfer function in the BDA algorithm. In contrast to previous binary metaheuristics, BDA allows the dragonflies to select values other than 1 and 0 using this transfer function. To obtain promising search space, BDA has a high exploration capacity. The reliable features of skin dermoscopy images are effectively selected by BDA based on the explanations given above.

Classification

Based on the selected features, the multi-class classification algorithm is used in the proposed MSCDNet model for effective image classification. The multi-class skin cancer classification uses a deep learning-based SqueezeNet model. During optimization, the SqueezeNet convolutional network is used. A test set is an input used in testing. The training and optimization stages for classification are done by the SqueezeNet convolution network. The layers of the fire modules are squeezed and expanded to create a simpler and more effective CNN model. The classification performance of the infection classes is determined by this most effective network model. The SqueezeNet model structure is given in Fig. 4.

The fifteen layers are presented in the SqueezeNet structure, including one global average pooling layer, three max-pooling layers, two convolution layers, one softmax output layer, and eight fire layers. K × K notation denotes the filter’s receptive field size, the feature map length is represented by l, and the stride size is denoted by s. The features of the input to the network are RGB channels, and the input image size is 227 × 227. The convolution is used for normalizing the input images and applying the maximum pooling to the input images. In the input volumes, the convolution layer uses 3 × 3 kernels to convene small regions and weights.

An element-wise activation function is performed by each convolution layer, which is the conclusive part of its argument. SqueezeNet, consisting of squeezing and expansion stages, uses fire in its convolution layers. An identical input and output tensor scale exist for the fire. For classification, the 1 × 1 size filter is utilized by the squeeze phase, whereas the expansion phase uses the 3 × 3 and 1 × 1 size filters. The squeeze initially transmits the H × W × C input tensor. Also, the C/4 of the input tensor channels is equal to the number of convolutions. The data depth is increased by C/2 of the output tensor depth through the expansions following the initial phase of transmission of the information. The ReLu units are employed throughout the extension and squeezing processes. When keeping the feature’s size the same, the expansion and squeeze operations compress and expand the depth. Concatenate action is used for organizing the expansion outputs in the depth dimension of the input tensor. The feature maps are described as FM and channels are represented as C, a layer of output from the squeeze process is described as f{y}, and the kernel can be used to represent w, i.e., (10)${\displaystyle f\left\{y\right\}=\sum _{fm1=1}^{FM}\sum _{c=1}^c{w}_c^f{x}_c^{fm1}}$

Here, f{y} ∈ R^N and w ∈ R^c×1×FM2. Using the squeeze outputs, a weighted combination of several feature maps of a tensor is established. A downsampling operation in the spatial dimensions is performed by a downsampling operation in the network. The network uses an average global pool for creating a single value from the class maps with features. At the network’s result, multi-class probability distributions are created by utilizing the softmax activation function.

Segmentation

A cascaded DenseUnet network model is developed to segment skin lesions precisely. For semantic segmentation, we carefully add the UNet model to the densenet model. The structure of DenseUnet is shown in Fig. 5. The down-sampling and up-sampling paths are used by this segmentation network for implementing the design concept of encoder–decoder of u-net and FCN. The continuous convolution operations are used for high-level and low-level semantic information extraction for the input images through the downsampling path.

Until the complete resolution of the input images is recovered, the resolution of the feature map is expanded to the upsampling path through deconvolution. There are four transition layers and four dense blocks in the upsampling and downsampling paths. n consecutive convolution layers of the exact resolution are used to build the dense block. Each layer builds the dense block using a dropout layer, rectified linear units (ReLU), and batch normalization (BN).

The output feature maps are obtained by each layer’s input, specifically from all of its previous layers, which is called dense concatenation. The dense blocks must have the same feature size in each layer’s feature maps for the dense concatenation operation. There are 10 dense blocks in the proposed Dense-UNet model, such as five dense blocks in the dense upsampling path and five dense blocks in the dense downsampling path. Four densely connected layers are used for forming each dense block; each layer has the same feature size.

Layer transition is achieved by providing the transition block. The dense downsampling path consists of five layers, each of which is made up of a transition block and a dense block. In the dense upsampling path, the high-resolution images are reconstructed using dense block operation, merger operation, and upsampling layer. The five layers are used for making the path. The regions are localized, and the complete input resolution is recovered by this path.

In a convolutional neural network, the lth layer output is described by x_l, and from the output of the preceding layer, a transformation Q_l(x) can be used to calculate x_l x_l−1 as: (11)${\displaystyle x_l=Q_l\left(x_{l-1}\right)}$ where Q_l(x) is a sequence of operations, where the sequence of operations is depicted as Q_l(x) convolution, pooling, rectified linear unit (ReLU), batch normalization (BN). From the previous layer, the skip connection output Q_l(x) with the feature maps is integrated to improve the flow of information and enhance the training against vanishing gradients: (12)${\displaystyle x_l=Q\left(x_{l-1}\right)+x_{l-1}.}$

However, the summation function is used for combining the output of Q_l an identity function, which can obstruct the flow of information in the network.

The skip connection is extended by introducing the dense connection to increase the information flow in the network. x_l isparticularly specified as: (13)${\displaystyle x_l=Q_l\left(\left[x_{l-1},x_{l-2},\dots ,x_0\right]\right)}$

Where the concatenation operation is represented by […]. A nonlinear transformation function is represented by H, which consists of dropout, Relu, BN, and convolution in this case. A transition layer connects the two dense blocks. A subsequent 2 * 2 average pooling layer, 1 * 1 convolutional layer, and BN layer are used to form the transition layer in the segmentation network. The flowchart of the proposed algorithm is shown in Fig. 6.

Evaluation measures

Resolved in this research are the problems with multi-classification. Melanoma, melanocytic nevi, benign keratosis, and basal cell carcinoma are accurately classified. Using a confusion matrix, an evaluation of the model’s effectiveness was conducted. F1-score, recall, precision, and accuracy are utilized for analyzing the proposed MSCDNet model performance.

The model is effectively classified by associating the model with a particular class, and the true positive (TP) indices comprise all data samples. In the confusion matrix, the true negative (TN) indices have additional samples. These samples are related to other successfully determined classes. The classifier’s estimated number of incorrect samples is referred to with the false negative (FN) and false positive (FP) indices in the uncertainty matrix. The performance metrics of the proposed model are calculated by using the following Eqs. (7)–(10).

In Eq. (1), accuracy is the proportion of true negative and true positive values divided by the confusion matrix’s total values. In the proposed model, the performance of the classifier is determined depending on the confusion matrix’s factors. (14)${\displaystyle Recall=\frac{TP}{TP+FN}}$ (15)${\displaystyle Precision=\frac{TP}{TP+FP}}$ (16)${\displaystyle Accuracy\left(\text{\%}\right)=\frac{TP+TN}{TP+FN+TN+FP}}$ (17)${\displaystyle F1-Score=\frac{2\left(TP\right)}{2\left(TP\right)+FP+FN}}$

The computation of the evaluation metric is done by using ground truth labels and predicted labels.

Experimental results

The dataset images are divided into training, validation, and testing for experiment analysis. As training and validation sets, multiple types of malignant images are used for training the proposed MSCDNet model. A total of 50 epochs are used in the experiment. In training and validation, the proposed model achieved the predicted accuracy once all of the epochs were finished. The prominent features are combined and then classified images into the multi-classification of skin cancers by the proposed MSCDNet model for constructing the output images. For multiple classifications of skin cancer diseases, multiple performance measures are used for evaluating the proposed MSCDNet model. The dermoscopic images are used for assessing the proposed MSCDNet model’s classification accuracy. The training data of each class is used for training the proposed model. For testing the proposed model, available test data is used. The results of the suggested model’s experiments are provided in Fig. 7

The proposed model’s performance metrics are extracted by the confusion matrices, which are displayed in Table 2. The class-wise F1-score, recall, precision, and accuracy of the proposed model are shown in this table.

Melanoma (MEL) is classified with an accuracy of 98.54%, 98.87% for benign keratosis (BKL), 99.01% for melanocytic nevi (MNV), and 98.67% for basal cell carcinoma (BCC) classification. The performance results for the classification of the proposed model are effectively enhanced from the experimental results. The proposed model produced 98.77% accuracy, 98.76% F1-score, 98.42% recall, and 98.56% precision. The proposed model’s performance in terms of classes is determined by using a representative subset of the validation set’s images. The suggested deep learning model efficiently handles several skin cancer disorders of different classifications.

The ROC curve and AUC score for each class are displayed in Fig. 8. All classes in the model have excellent AUC scores. If the model achieved the maximum AUC of the ROC, it is considered appropriate and effective. The ROC curve is calculated using the false-positive and true-positive rates. An AUC (ROC) of 0.9516 is attained by the proposed model.

Figure 9 displays the MSCDNet validation and training accuracy throughout 50 epochs. A maximum of 98.77% accuracy is obtained for training, and a maximum of 92.15% accuracy is obtained for validation from the experimental results. Both training and validation achieve the highest level of performance. The model attains 0.011 for training and 0.069 for validation. The complete training process uses a timing of 1 min 54 s to arrive at the final iteration. The proposed MSCDNet is adequately trained and can correctly classify multiple skin malignancies. The proposed model’s confusion matrix is displayed in Fig. 10.

Comparison results

The classification results of the proposed model with those of earlier studies are compared that used the same dataset in this section. With the most existing classifiers currently available, as listed in Table 3, the proposed MSCDNet is thoroughly analyzed regarding precision, f1 score, recall, and accuracy.

Inception v3-based Convnet net model for skin cancer diagnosis is provided by Mijwil (2021). The proposed model analyzes skin cancer into benign and malignant types. For multiple classifications of skin lesions, the Weighted Average Ensemble classifier is developed by Rahman et al. (2021), and this approach achieves an 88.00% of accuracy rate. 96.30% accuracy is achieved by Khan et al. (2021a) for multi-class skin cancer classification. A 92.83% accuracy rate for the multiple classifications of skin cancer is attained by Chaturvedi, Tembhurne & Diwan (2020). While comparing to Iqbal et al. (2021) and Zhang et al. (2019), 89.58% accuracy is achieved by Iqbal et al. (2021) and 86.80% accuracy is achieved by Zhang et al. (2019) for binary skin cancer classification. Kassem, Hosny & Fouad (2020) proposed a modified GoogleNet model for multi-skin lesions classification, achieving an 81% accuracy. The proposed MSCDNet classification model performs better when compared to the previous deep learning models and successfully detects the multi-class skin lesion classification. Due to the classification process, the overfitting and adverse effects on network performance are prevented by using the proposed classification network model. The comparative analysis of accuracy for the state-of-the-art model is shown in Fig. 11.

The inter-scale variability of the disease features is successfully extracted by using the proposed model. Additionally, the classification results of the proposed model are increased. The proposed method performs better than the existing methods for multiple skin cancer classification achieving 98.77% F1-score, 98.42% recall, 98.56% precision, and 98.77% accuracy. The suggested model’s impressive classification accuracy enables automated diagnosis and pre-screening of brain tumors.