Banca de DEFESA: PEDRO VITOR SOARES GOMES DE LIMA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : PEDRO VITOR SOARES GOMES DE LIMA
DATE: 30/03/2023
TIME: 14:00
LOCAL: Remota
TITLE:

Intelligent system to support the diagnosis of dermatological diseases using digital image analysis and deep artificial neural networks


KEY WORDS:

Diagnostic support system. Leprosy. Skin cancer. Convolutional neural network.


PAGES: 99
BIG AREA: Engenharias
AREA: Engenharia Biomédica
SUMMARY:
According to the World Health Organization (WHO), neglected diseases are a set of infectious diseases that are common in low-income populations in developing regions. They do not have sufficient appeal to receive government or market investments, either due to low prevalence or simply because those are diseases of the poorer sections of the population. One of the diseases classified as neglected is leprosy, with Brazil being one of the countries with the most cases, reaching an overall new cases detection rate of 13.23 per 100,000 inhabitants in 2019. With a different reality, but also worrying, cancer is considered one of the leading causes of death and a critical barrier to increasing world life expectancy. In Brazil, for each year of the triennium 2020-2022, there will be an estimated 625 thousand new cancer cases, with non-melanoma skin cancer being the one with the highest incidence (177 thousand). Knowing this fact, we aim to create a solution that contributes to the early detection of these diseases. A diagnostic support system is proposed for skin diseases, more specifically: for leprosy and skin cancer. By using deep learning for feature extraction combined with different classification techniques, the intelligent system can recognize diseases by analyzing skin lesions pictures. Using Weka software libraries, we analyzed the convolutional neural networks LeNet, SqueezeNet, NASNetMobile, ResNet50, and Inception V3, combined with the Random Forest classifier with 50, 100, 150, 200, 250, and 300 trees. From boxplots and tables with mean and standard deviation, we evaluated the following metrics: accuracy, specificity, sensitivity, area under ROC curve, and kappa index. We observe that the convolutional neural network Inception V3 is the one that usually generates better results and that, considering a trade-off between performance and computational cost, the appropriate configuration for the classifier is the one with 150 trees. Thus, reaching an accuracy of88.59±1.93%, specificity of97.30±1.12%, sensitivity of 71.00±7.39%, area under ROC curve of94.80±1.78%, and kappa index of86.69±2.25%, considering the classification among seven diseases (four types of leprosy and three types of skin cancer).

COMMITTEE MEMBERS:
Presidente - 1807632 - WELLINGTON PINHEIRO DOS SANTOS
Interno - 1170977 - RICARDO EMMANUEL DE SOUZA
Externa à Instituição - ELLANY GURGEL COSME DO NASCIMENTO - UERN
Notícia cadastrada em: 08/03/2023 09:44
SIGAA | Superintendência de Tecnologia da Informação (STI-UFPE) - (81) 2126-7777 | Copyright © 2006-2024 - UFRN - sigaa06.ufpe.br.sigaa06