Desenvolvimento de uma gramática para construção automática de arquiteturas de deep learning aplicado à classificação de imagens dermatoscópicas

Abstract

Melanoma is a type of cancer originated in the pigment-producing melanocytes of the skin and caused by over production of melanin pigments by melanocytes. It is also known as malignant melanoma and often resembles moles or pigmented skin lesions. Melanoma is the most dangerous type of skin cancer. The continuous increase in incidence rates and melanoma mortality have guided searchs aiming to help specially on early detections, where physicists usually fail on detecting and the diagnosis is most efficient. Therefore, the use of computational tools helping the analysis of skin lesions is very important. These tools can increase the accuracy and precision of the correct diagnosis of the lesion. Another important aspect that should be considered is the democratization of the disease diagnosis sharing knowledge that used to be limited to specialized people. Deep Learning is a research area under the spotlight in recent years due to its successful application to many domains, such as computer vision and image recognition. The most prominent technique derived from Deep Learning is Convolutional Neural Network, which allows the network to learn representations needed for detection or classification tasks automatically. Deep learning convolutional neural networks (CNN) may facilitate melanoma detection. However, CNN have some limitations, as designing these networks are not easy to master and require expertise and insight. In this work, we present the use of genetic algorithm associated to grammar-based genetic programming to optimize convolution neural network architectures. It is presented a grammar capable of exploring a search space from the problem. The search using the grammar ensure the criation of valid individuos that can be qualitatively analized. To evaluate our proposed approach, we adopted images from ISIC dataset to validate the evolution of the generated architectures, using the metric of accuracy to evaluate its classification performance in the test dataset. The results demonstrate that our method using grammar-based genetic programming can easily produce optimized CNN architectures that are competitive and achieve high accuracy results, showing superior results in comparison with the state-of-the-art methods.