Classification Model of Skin Cancer Using Convolutional Neural Network

Cancer is a worldwide health problem, not only because of its high incidence and mortality rates, but also because of the high social cost it represents for each country. The General Directorate of Epidemiology (DGE) of Peru has conducted an investigation of the cancer statistics in this country based on the epidemiological surveillance of cancer, finding that in the period 2006 – 2010, with 5975 cases of skin cancer registered, representing 6.6% of all cancers registered. According to this report, skin cancer ranks fourth in frequency at the national level [1].

Furthermore, in the period 2014 - 2018 the most frequent types of cancer globally were those of the cervix (18.6%), stomach (11.1%) and skin (10.8%) [2]. Skin cancer has been increasing in recent decades, mainly due to population aging and exposure to ultraviolet (UV) radiation, changes in human behavior and ozone depletion [3]. According to the International Agency for Research on Cancer (IARC) in 2022 there were more than 1.5 million new cases of cancer and approximately 330,000 cases of melanoma cancer, the most dangerous type of skin cancer [4].

The skin cancer was the most frequently diagnosed cancer in Metropolitan Lima, also in several coastal regions it is the second in frequency [2].

Three categories of skin cancer include “basal cell carcinoma”, “squamous cell carcinoma”, and “melanoma” [5], however, in our article, we classified skin lesions into 9 types, which will be described later.

Traditional methods of skin cancer detection, which include visual inspection and biopsies, have problems with regard to accuracy and accessibility, relying heavily on the expertise of the dermatologist which poses an obstacle to many regions where there is little access to trained specialists. Although machine learning has emerged as an alternative detection method, it faces limitations related to the diversity of data used to train it, as well as acceptance by healthcare professionals and patients [6].

This article outlines the creation and structure of an AI model employing convolutional neural networks for analyzing skin medical images. Its purpose is to identify and classify cancer types among those mentioned based on the detected abnormalities. Deep neural networks demonstrate remarkable capabilities in facilitating efficient planning optimization and proficiently addressing a diverse spectrum of problem-solving scenarios [7-9].

The most important factor for the success of an application is the optimal organization of its data [10-16].

3.1 ISIC dataset

The “International Skin Imaging Collaboration” (ISIC) have a public dataset (2019), to comprises dermoscopic images with “normal” and “skin cancer”. It was created to gather skin lesion images under Creative Commons licenses, featuring reference diagnoses and additional clinical metadata for comprehensive use [31]. This dataset was used for the SIIM-ISIC 2020 melanoma classification challenge, organized in Kaggle during the summer of 2020.

The Kaggle dataset of 2019 contains 2357 dermoscopic images of different skin lesions classified into 9 types [32]:

“Actinic keratosis” (130 images)

“Basal cell carcinoma” (392 images)

“Dermatofibroma” (111 images)

“Melanoma” (454 images)

“Nevus” (373 images)

“Pigmented benign keratosis” (478 images)

“Seborrheic keratosis” (80 images)

“Squamous cell carcinoma” (197 images)

“Vascular lesion” (142 images)

This dataset presents an imbalance in the number of samples per class, which may bias the model towards those classes with more images, such as “Pigmented benign keratosis” or “Basal cell carcinoma”, and affect the performance in less represented classes, such as “Seborrheic keratosis” or “Dermatofibroma”. To mitigate this problem, it was deemed necessary to apply data augmentation techniques, which generate new instances and balance the distribution of classes. This improves the ability of the model to generalize and correctly recognize lesions across categories [33]. Then, the size of the images was then set to 180×180 pixels, as this is large enough to capture important details, but not so large that the training would be too slow.

3.2 Pre-processing

The ISIC dataset provides a diverse collection of dermatoscopic images, classifying skin lesions into nine distinct types. Each type represents a specific pathology, allowing for a comprehensive analysis of various dermatologic conditions.

To improve the performance of the model in the classification of these images, it was considered necessary to apply data augmentation techniques. These techniques are essential when working with reduced data sets, since they allow the generation of new samples from the original images, thus increasing the volume of available data. In this way, the distribution of classes is balanced, thus improving the accuracy and robustness of the model [33].

In this project, rotations, cropping and scaling were implemented as key transformations to increase image variability. Image rotation allows the model to learn to recognize skin lesions in different orientations, which is critical in medical image analysis where lesions can appear in various positions. Cropping helps the model to recognize key features of lesions, even when they are not perfectly centered or when parts of the image are not visible. On the other hand, scaling allows the model to identify lesions of different sizes, simulating variations in distance or field of view, which is particularly useful when using images captured with different devices or under varying conditions.

Previous studies have shown that these techniques improve the model's ability to correctly identify malignant lesions [5, 22, 25].

Therefore, these data augmentation techniques have proven to be effective in improving the generalization capacity of the model in image classification tasks. Thanks to these transformations, we were able to increase the dataset to 1000 images per class, reaching a total of 9000 images. This not only improves the distribution of classes, but also provides a greater diversity of images to train the model, making it more robust and reliable in the classification task.

3.3 Training and testing strategy

In this project, the training and testing split was 80 and 20 respectively of the developed classification modules, based on the assumption of "Pareto's Law" referred by Bustamante [34], where the small proportion (20%) are vital while the remaining (80%) are trivial [19].

As mentioned, rotations, cropping and scaling were applied as transformations, resulting in an increase to 1000 images for each class. The result was a total of 9000 images.

Table 2 presents the proposed distribution for the dataset of ISIC and Table 3 shows the segmentation of the images after data augmentation.

Table 2. Images processed per operation (training and testing) in each instance [19]

Table 3. Quantity of images per operation post data augmentation [19]

3.4 Design and development

In designing the convolutional neural network (CNN) model for this study, we followed an architecture inspired by previous work on skin cancer classification using deep learning techniques. Our CNN includes an input layer for image data, followed by multiple normalization and convolutional layers, each accompanied by clustering layers to reduce dimensionality, and concluded with two fully connected layers. The first dense layer uses ReLU activation, and the final layer applies softmax activation for classification. This architecture is similar to those used by some researches [23, 26, 27], who have successfully applied CNN models for medical image classification tasks, demonstrating the effectiveness of deep learning in skin cancer detection using image analysis.

The graph provided by Figure 1 details the development flow of the proposed convolutional neural network (CNN) model. This sequence spans from the initial stages of data collection to the final phase of model analysis.

1.jpg

Figure 1. CNN development process

The procedure initiates by identifying data for model training and testing, followed by data augmentation and preparation. Following the steps of another study [7], it is divided into validation and training sets, organized into folders for each class, and efficiently managed in a Google Drive repository.

As we advance in the diagram, we observe the configuration of the layers of the convolutional neural network (Table 4).

Table 4. Layers of the sequential model [19]

The structure consists of 19 layers, with the initial Rescaling layer normalizing image pixel values to the [0, 1] range and the images are 180 × 180 pixels and have 3 color channels (RGB) as can be seen in Table 4.

The network consists of convolution layers, each with ReLU activation, and pooling layers utilizing a 2 × 2 pooling window. This pairing (convolution layer – pooling layer) is repeated six times. Filter counts increase progressively from 16 in the first convolution layer to 512 in the sixth. Following the last three convolution layers, dropout layers are applied for regularization. Following is the flattening layer, which transforms the output into a one-dimensional vector for readiness for the dense layers. Next, two consecutive dense layers are employed: the initial layer shows 1024 neurons with ReLU activation, and the subsequent layer shows 9 neurons, representing the number of classes, with a softmax activation function. This function transforms outputs into probabilities, assigning a likelihood to each class, with the highest probability determining the model's prediction [35].

The convolutional neural network will use optimizers that allow improving its prediction level, as shown in Table 5.

Table 5. Optimizer setting

Table 6. Parameters setting for the training

The next stage of the figure shows us the training of the model. Table 6 details the parameters for training this model. Where the train and validation data are those found in the validation and training set resulting from the division of the set of images after the data augmentation. Training is performed for 25 epochs, and in each epoch, a number of steps are taken equal to the length of the training and validation data sets, respectively.

Finally, the last stage of the figure refers to the analysis of the model where the training and validation accuracy and loss are analyzed. To improve visualization, line graphs were used.

Just as other study [7], the development of the model was done in Google Colab, using Python 3, keras of Tensorflow, numpy, matplotlib, pandas, seaborn and Data augmentation libraries.

When compared with other studies, our results are in agreement with those of previous studies. For example, Benbrahim et al. [26] achieved a similar accuracy of 94.06% using a CNN model trained on the HAM10000 dataset, but their recall was slightly lower, suggesting that our model is more sensitive in identifying positive cases. If we compare with the study [37], which obtained accuracies of 92.30%, 93.95% and 94.31% on ISIC 2019, HAM10000 and a private dataset, respectively, our model performs better on ISIC 2019 and HAM10000 datasets. However, the strength of our model lies in its specificity, which, at 98.69%, exceeds that of many other studies. This indicates that our model is particularly effective in minimizing false positives, which is crucial in clinical settings to avoid unnecessary treatments.

Additionally, when compared to the results of Kassem et al. [38], who achieved a classification accuracy of 94.92% with a sensitivity of 79.8% and specificity of 97% using transfer learning with GoogleNet, our model demonstrates superior recall (89.55%) and specificity (98.69%).

While Kassem’s model shows a strong accuracy, the relatively lower recall and precision (80.36%) indicate that our model might offer better performance in detecting true positives and reducing false negatives. This balance between sensitivity and specificity is essential, particularly in medical diagnostics where both false positives and negatives can have significant consequences.

When comparing the results obtained with Cauvery et al. [39], who reported a balanced accuracy of 81.2%, while our model achieved an overall accuracy (Accuracy) of 94%, which represents a remarkable difference in terms of overall accuracy.

Furthermore, our precision of 89.88% exceeds that of the model of Cauvery et al. [39] who reported an average accuracy of 73% in their best ensemble. This suggests that our approach more effectively handles the reduction of false positives. In terms of sensitivity (Recall), our model achieved 89.55%, improving the ability to detect malignant lesions compared to 62% in the model of Cauvery et al. [39], likewise, the F1 score obtained in our work (89.4%) suggests a better balance between accuracy and sensitivity, surpassing the 61% reported by them. These results reflect that our implementation offers superior performance in the detection of skin cancers.