DYNA JOURNAL ENGINEERING

Key words:

skin cancer, CNN models, HAM10000, classification, dataset

Article type:

ARTICULO DE INVESTIGACION / RESEARCH ARTICLE

Section:

RESEARCH ARTICLES

Early detection of skin cancer is quite important since some types, such as melanoma, are dangerous and even it can cause the death if are not properly and early treated. Here, deep learning is broadly used to implement non-invasive systems for diagnosing skin cancer based on image analysis. In this work, it is presented a model, which is a derivation of the Visual Geometry Group with 16-layer deep model architecture (VGG16), for implementing a benign keratosis (bkl), melanocytic nevi (nv) and melanoma (mel) skin cancer classifier. Moreover, it is shown that by using a balanced dataset an average classification accuracy of 79.94% for the three types of cancer can be reached. Furthermore, it is presented this accuracy is quite competitive compared when the same skin cancer classifier is implemented by using the VGG16, the ResNet50v2, and the InceptionV3 pre-trained models, since the obtained average accuracies were 78.32%, 79.40% and 81.03%, respectively, under the same conditions. Additionally, it is described that the proposed model is lighter compared with the mentioned pre-trained models since it requires of 1.4 million of trainable parameters. Finally, it is shown that this characteristic contributes to the computational processing time for implementing, training, and evaluating the skin cancer classifier when the proposed model is used. Based on these numerical results, it is shown that the proposed model is competitive in terms of the required number of trainable parameters and the overall processing time in comparison with the required by the mentioned pretrained models without penalizing the skin cancer classification metrics.

Keywords: skin cancer, CNN models, HAM10000, classification, dataset.