Accurate Deep Learning Algorithms for Skin Lesion Classification

Skin cancer, is a disease that mainly affects men and women equally; however, it is somewhat more common in men. Skin cancer is responsible for 4% of cancer cases in women and 6% of cancer cases in men [1]. It is a major public health concern owing to its rising prevalence and potential severity. Factors leading to the growth in skin cancer cases are [2]:

UV Exposure: Prolonged exposure to ultraviolet (UV) light from the sun or tanning beds increases the chance of developing all forms of skin cancer.

Genetic Predisposition: People who have specific genetic conditions or a family history of skin cancer are more susceptible.

Skin Type: Fair skinned, light-haired, and light-eyed individuals are more vulnerable to UV exposure and ensuing skin cancer.

Immune Suppression: Those who have compromised immune systems—from diseases or medical procedures—are more vulnerable.

Skin lesions refer to a wide range of disorders, ranging from harmless to cancerous, each having unique features and consequences for health. Benign lesions include melanocytic nevi, or moles, as they are more frequently called, which are mostly innocuous but may sometimes undergo a transformation into melanoma. Importantly, melanoma is a type of skin cancer that can be fatal and arises from melanocytes, which are cells that produce color [3]. Past sunburns and excessive sun exposure frequently contribute to its development. For effective melanoma treatment and better patient outcomes, early identification and diagnosis are essential. The information highlights how critical it is to raise awareness, practice prevention, and conduct routine skin exams in order to detect melanoma in its early stages, when it is most curable. Other benign lesions, like dermatofibromas and seborrheic keratoses, typically pose no significant risk, but their removal may occur if they begin to cause symptoms or for aesthetic reasons. Actinic keratoses are precancerous lesions that, if left untreated, may develop into squamous cell carcinoma (SCC). This emphasizes the need for early discovery and treatment. Basal cell carcinoma (BCC), is the most prevalent form of skin cancer in White communities, accounting for 75% of all skin malignancies. Typically, they appear in the head and neck region (w83%) [4]. Squamous cell carcinoma (SCC) The incidence of this type is lower than that of BCC. It grows in arid and rugged regions and between squamous cells. Nevertheless, it may manifest in specific regions of the epidermis that receive an increased amount of radiation. It may manifest as red patches. Similar to BCC, it has the potential to disseminate to other parts of the body; however, certain treatments have been identified to avert its progression [5]. Vascular lesions, such as angiomas and pyogenic granulomas, are noncancerous growths of blood vessels that often appear as red or purple patches on the skin [6].

The artificial neural network is utilized by deep learning to examine data in various categories and uncover patterns to learn from it. This type of learning is based on artificial neural network architecture, which they are built to emulate the structural design of the human brain. These convolutions, are in fact, layers that serve to identify and filter only pertinent information from the data [7].

Deep learning mainly uses artificial neural networks to analyze images of skin disease and detect skin cancer. It has great predictive ability in the diagnosis of skin cancer with a primary role in most malignant cancers [8]. Deep learning, in turn, is a basis of training of artificial neural networks with large data sets.

The challenges facing skin cancer diagnosis techniques are numerous, including the few annotated datasets and the need of significant processing power. Nevertheless, the effectiveness of deep learning relies on process development that has the potential to resolve these problems and improve the diagnostic precision of skin cancer.

Skin lesions are particularly interesting and useful in the context of CNNs to identify and classify the lesions [9]. Convolutional neural networks (CNNs) are a type of neural network that are specifically designed for image recognition tasks, and of course take advantage of hierarchial structure that are supposed to abstract important features from the images. CNN is a composition of many layers collaborate with each other in features extraction and classification. The beginning layers of the architecture of a neural network identify basic features of an image such as edges & corners, while the latest layers will pick-up more complex features. E.g. shapes, patterns. The last layer of the network outputs the classification result, that is, the type of skin lesion. This allows the network to learn increasingly complex representations of the input data as it goes through the layers of the network.

CNNs are quite good at correctly categorizing skin lesions—more especially, they are very good at melanoma detection. Unfortunately, there are frequently few annotated datasets available in the dermatological sector, which poses a serious problem because CNNs need a sizable amount of labeled data to train [10].

The application of transfer learning in deep learning is a tactic meant to increase model performance on fewer datasets. By utilizing the knowledge gathered from previously trained models on larger datasets, this is accomplished. This way of classifying skin lesions uses CNN models that have already been trained on large datasets (ImageNet) to pull out important features from images of skin lesions. A smaller CNN model is then trained using these extracted features on a more compact series of dataset of skin lesions [11].

It has been demonstrated that transfer learning increases skin lesion categorization accuracy, especially when there is little labeled data available. However, the pre-trained model selection and the particular architecture of the smaller CNN can have a significant impact on the transfer learning performance.

In spite of the fact that diagnostic technologies have seen some advancements, there are still several obstacles and constraints that remain in the area of skin lesion classification:

Subjectivity and Variability in Visual Inspections: Conventional approaches to identifying skin lesions mainly depend on visual inspections conducted by dermatologists. The subjectivity and variability about the matter may yield inconsistent diagnoses, misdiagnoses and ultimately late-stage delayed treatment.

Data Imbalance: The lack of an even distribution between benign lesion and malignant lesions, makes the varied appearance of skin lesions a real challenge. Such potential outcome could be adversely biased the functioning of a model.

Inadequate Annotated Data: Datasets with high-quality annotated data are very important for the well-fitting of deep learning models. Provided these datasets house the necessary ground truth for supervised learning algorithms to correctly understand and classify various skin lesions. But the scarcity of available information makes it particularly difficult for the medical field to develop and validate reliable diagnosis models.

In this paper, We investigate and conduct a thorough analysis of different deep learning algorithms for skin lesion detection, such as Convolutional Neural Networks and transfer learning models, like DenseNet201 and ResNet52V2. Our results demonstrate that the proposed models achieving 95% accuracy on training images and 91% accuracy on test image, pointing their ability to appear in classifying skin lesion types. This study aims to aid ongoing efforts to improve skin cancer detection and outcomes of patients with Artificial Intelligence technology.

3.1 HAM10000 dataset

The HAM10000 [19] is a large dataset of skin lesions acquired for a scientific research purpose from the Department of Dermatology at the Medical University, Vienna. It consists of 10,015 dermoscopic images categorized into seven different groups showing pigmented skin lesions:

(1) (AKIEC) Actinic keratosis / Bowen's disease. the lesion is a non-malignant tumor, but that may evolve into a malignant disease (squamous cell carcinoma).

(2) Malignant - (BCC) Basal cell carcinoma.

(3) Non-malignant – (BKL) keratosis Lesion.

(4) Benign – (DF) Dermatofibroma.

(5) Malignant – (MEL) Melanoma.

(6) Benign (non-cancerous) - (NV) Nevi Melanocytic.

(7) (VASC) Vascular lesions: can be either Malignant or Non-malignant lesion.

Figure 1 shows several representative images of skin lesions. The collection consists of images with varying resolutions. Because to resolution variability, categorization of skin lesions becomes challanging task.

3.2 Pre-process methodology

To improve the quality of images and ensure their compatibility with the requirements of neural networks used in classification and pattern recognition. The dataset is organized into six folders and known as class names. Because the raw dataset is highly imbalanced, data balancing was performed using down sampling method. The dataset is then divided into two sets: 20% for testing and 80% for training. The training and test sets were then subjected to image preparation methods, which included image normalization the augmentation. Figure 2. illustrate the main phases of preprocessing method.

1.png

Figure 1. Sample photos from the dataset

Preprocessing of HAM10000 dataset images was applied in the following steps

First, resizing images to fixed dimensions is a crucial step to feed the neural network. All images in the dataset were resized to 224 × 224 pixels, which is the optimal size for many pre-trained neural networks such as ResNet and DenseNet. This change ensures that all images enter the network with the same dimensions, reducing the complexity of the learning process and increasing training efficiency [20].

Second, image balancing: The dataset is highly imbalanced, with the (Melanocytic nevi) class representing the majority of the images (31.2%), and the other classes having much fewer samples (akiec: 6.8%, bcc: 10.7%, bkl: 22.8%, df: 2.4%, mel: 23.1%, vasc: 3.0%) shown in Figure 3. The images are of various sizes and resolutions, and they were acquired from a range of sources and conditions, resulting in variability in terms of lighting, focus, and quality.

To address class imbalance in datasets, one way is to resample the training dataset randomly. Undersampling entails removing instances from the majority class, and oversampling [21] entails copying examples from the minority class. Three phases made up the oversampling technique utilized in this study:

1) First, a subset of the HAM10000 dataset that contains only the "Melanocytic nevi" class is created.

2) The subset is then randomly sampled to reduce the number of samples in the "Melanocytic nevi" class to 1500.

3) Next, the remaining classes in the HAM10000 dataset are oversampled to balance the number of samples across all classes. This can be achieved using techniques such as random oversampling, which involves randomly duplicating samples from the minority classes until the number of samples in all classes is equal. The result is shown in Figure 4.

2.png

Figure 2. Preprocessing phases

3.png

Figure 3. Database before balancing

Third, normalizing (adjusting the pixel values to range from 0 to 1): adjusting is done by dividing the value of each pixel by 255 (the maximum pixel value in grayscale images or color images). This uniform range makes it easier for the neural network to process images better and ensures the stability of the input values [22, 23].

Fourth: image augmentation: When the number of images has decreased due to the use of image balancing, a technique is used to increase the diversity of images available to train deep learning models without actually collecting new images. This was achieved by applying different transformations to the HAM10000 dataset images, such as rotating, zooming in and out, flipping horizontally and vertically [24].

By performing these four steps: Image Resizing, Image Balancing, Pixel Scale Adjustment, and, Image Augmentation which speeds up the neural network's training process and contributes to improve the model’s performance and increasing the classification accuracy.

4.png

Figure 4. Database after balancing

3.3 DenseNet201

DenseNet201 is a convolutional neural network architecture that was introduced in 2018 as an extension to the original DenseNet architecture. DenseNet201 gets its name from the sum of 201 layers that it contains [25].

DenseNet201, like the original DenseNet architecture, makes use of a densely connected network structure in which all layers above it provide input to the layer below it, encouraging feature reuse and lowering the number of parameters.

By adding bottleneck layers this reduces the computing cost of the network. They use 1x1 convolutions to reduce the number of input channels before the main convolution operation is performed. They then use another 1x1 convolution to go back to the original output channels. Meaning, the network works faster and more efficient and needs to learn fewer parameters during the convolution process.

Moreover, DenseNet201 utilizes feature concatenation which concatenates feature maps from numerous layers before they proceed to the next layer. This is because the network has higher precision and a better feature acceptance mode for low-level and high-level features.

Overall, still DenseNet121 is the best deep learning framework for image classification tasks. It has high use in computer vision mainly used is in medical image analysis.

3.4 ResNet152V2

Is a type of deep convolutional neural network structure. It has shown great success in the field of image classification, such as the classification of skin lesions. This is an extension of the ResNet family of architecture which used residual blocks to train deeper networks to prevent vanishing gradient problem [26].

ResNet152V2 is composed of 152 layers, most of which are convolutional [27]. The model contains four stages in its architecture, with the count of residual blocks being different throughout them. First stage 7 residual layers, second stage 35 residual layers, third stage 8 residual layers, final stage 3 residual blocks.

Each block is composed of two or three convolutional layers followed by ReLU activations and batch normalization. The network can learn residual functions that can be used to modify the input features by adding the output of the convolutional layers to the input of the residual block. The residual blocks also include skip connections, which enable gradient flow through the network during backpropagation and prevent the vanishing gradient problem.

ResNet152V2 has global average pooling in addition to convolutional layers and residual blocks. This decreases the spatial dimensions of the feature maps to one value per feature map. After that, a fully connected layer processes the generated feature vector to provide the output that is needed to complete the classification task.

ResNet152V2 has been shown to be highly effective in skin lesion classification, achieving high accuracy on various datasets including the HAM10000 dataset.

3.5 Metrics

Figure 5 illustrates confusion matrix which is utilized to valuate binary classification models by representing True (1) or False (0) for the actual classes and Positive (1) or Negative (0) for the resulting or predicted classes.

5.png

Figure 5. Confusion matrix [28, 29]

True Positive (TP): The predictive model precisely estimates the positive class.

False Positive (FP): The predictive model inaccurately predicts the positive class (also called a Type II error).

True Negative (TN): The predictive model accurately predicts the negative class.

False Negative (FN): The predictive model inaccurately estimated the negative class (also called a Type II error) [30].

Several important measurements may be derived from the confusion matrix to assess the classification model's performance:

Accuracy:

$ Accuracy =\frac{T P+T N}{T P+T N+F P+F N}$                 (1)

Accuracy is the percentage of accurately predicted cases (both positive and negative) out of all instances.

Precision (Positive Predictive Value):

$ Precision =\frac{T P}{T P+E P}$                             (2)

Precision refers to the fraction of genuine positive predictions among all positive predictions produced by the model.

Recall (Sensitivity or True Positive Rate):

$ Recall=\frac{T P}{T P+E N}$                   (3)

Recall calculates the fraction of genuine positive cases among all actual positive instances.

Specificity (True Negative Rate):

$ Specificity =\frac{T N}{T N+F P}$                 (4)

Specificity refers to the fraction of real negative cases among all actual negative instances.

F1 Score:

$F1 Score =2 \times \frac{Presision \times Recall}{resision+ Recall}$                   (5)

The F1 score is the combined mean of precision and recall, resulting in a balance of the two measurements.

Receiver Operating Characteristic (ROC) Curve:

The ROC curve shows how a binary classifier's diagnostic performance changes when the discrimination threshold is adjusted. It compares the actual positive rate (sensitivity or recall) with the false positive rate (1-specificity).

3.6 Proposed model

The proposed model to classify skin lesions involved several key steps as it shown in Figure 6. In order to match the input picture dimensions for the selected models, Densenet201 and ResNet152V2, a preprocessing step was utilized to scale the images to a consistent size of 224 × 224 pixels. Oversampling was used to address the issue of class imbalance in the HAM10000 dataset. To equalize the distribution of samples across all classes, more members of the minority class have to be created.

Using the preprocessed and oversampled HAM10000 dataset, the Densenet201 model was used in the study's initial phase, and it produced an accuracy of 88%. Using a pre-trained Densenet121 model—which attained an accuracy of 86%—was the second phase. Using a pre-trained Resnet152V2 model, the third stage resulted in an accuracy rate of 88%.

To further improve accuracy, a combined analysis of the output from the top-performing models was done.

In order to optimize the performance of the suggested model, the outputs of the Densenet201 and Resnet152V2 models were integrated by combining them, and then the following layers were added:

1. The Global Average Pooling Layer processes the averages across the whole spatial dimension of the feature maps after receiving the combined model outputs. As a result, a global feature vector is produced, which highlights the salient characteristics of the input image.

2. The addition of the dense layer, which has 256 neurons and a ReLU activation function, enables the network to learn more intricate features from the input data.

3. Dropout 0.2: During training, this layer arbitrarily removes 20% of the activations from the preceding dense layer. This regularization method aids in avoiding overfitting.

4. Dense layer (64, activation='relu'): This layer adds another dense layer with 64 neurons and applies the ReLU activation function. This further allows the network to learn more complex features of the input data.

5. The second Dropout layer randomly drops out 20% of the activations from the previous Dense layer during training.

6. The final output layer of the model, the Dense layer with 7 neurons, is included, one for each class in the dataset. Over the classes, a probability distribution is produced using the softmax activation function. For the input image, the predicted class is determined by selecting the class with the highest probability.

We used the categorical cross-entropy loss and the Adam optimizer for training the model. The model was trained for 60 epochs with the batch size of 64. The model was checkpointed with the best model weights based on validation loss. In order to prevent overfitting and increase the variability of the training data, image augmentation like (rotation, zoom of the images, changing the width and heigh shift range, flipping) were used randomly to train the images during the training. Lastly, a series of metrics, including the F1-score, Precision, Recall, AUC-ROC, and Classification Accuracy, were used to evaluate the performance of the recommended model on the test set.