| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 13.41 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A psoríase é uma lesão dermatológica que se manifesta em diversas regiões do corpo. Seu
diagnóstico tardio pode gerar agravamento da doença em si, bem como das comorbidades
associadas a ela. Contudo, devido à grande diversidade de doenças dermatológicas
existentes, este diagnóstico pode ser complexo.
Neste contexto, o trabalho proposto apresenta um sistema computacional para auxiliar
o processo de diagnóstico da psoríase por meio de redes neurais convolucionais profundas
e uma aplicação para dispositivos móveis.
Para classificação foi criado um dataset composto por 752 imagens de pele humana,
utilizado como alimentação às 3 redes de classificação: MobileNetV2, Xception e InceptionResnetV2.
Uma etapa de segmentação semântica através de 3 arquiteturas diferentes
baseadas na topologia da rede U-net foi desenvolvida para avaliar a influência da segmentação
na classificação.
Os resultados atingidos apontam um bom potencial do sistema para classificação de
imagens de psoríase. Um conjunto de testes entre diferentes tipos de dataset e redes
neurais foi aplicado e apontou que, apesar do modelo de segmentação proposto ter atingido
mais de 97% de acurácia e uma pontuação DICE superior a 91%, proporcionando
uma segmentação relativamente boa, o dataset original proporciona maior acurácia de
classificação em relação ao dataset segmentado. Na classificação, o modelo MobileNetV2
apresentou uma acurácia maior que 97% com uma taxa de perda de aproximadamente
7%.
O aplicativo móvel desenvolvido é capaz de utilizar os modelos de classificação e segmentação
de forma simples, auxiliando o utilizador final no processo de análise da lesão dermatológica por meio de uma interface intuitiva.
Psoriasis is a dermatological lesion that manifests itself in several regions of the body. Its late diagnosis can worsen the disease itself, as well as the comorbidities associated with it, however, amid the great diversity of existing dermatological diseases, this diagnosis can be complex. This work proposes a computational system to assist the psoriasis diagnosis process through deep convolutional neural networks and an application for mobile devices. For classification, a dataset was created composed of 752 images of human skin, fed to 3 classification networks: MobileNetV2, Xception and InceptionResnetV2. A semantic segmentation step through 3 different architectures based on the topology of the U-net network was developed to assess the influence of segmentation on the classification. The results present a good potential of the system for the classification of psoriasis images. A set of tests between different types of datasets and neural networks was applied and pointed out that, although the proposed segmentation model reached more than 97% accuracy and a DICE score greater than 91%, providing a relatively good segmentation, the original dataset provides greater classification accuracy compared to the segmented dataset. In the classification, the MobileNetV2 model presented an accuracy greater than 97% with a loss rate of approximately 7%. The developed mobile application is able to use the classification and segmentation models in a simple way, assisting the end user in the process of analyzing the dermatological lesion through an intuitive interface.
Psoriasis is a dermatological lesion that manifests itself in several regions of the body. Its late diagnosis can worsen the disease itself, as well as the comorbidities associated with it, however, amid the great diversity of existing dermatological diseases, this diagnosis can be complex. This work proposes a computational system to assist the psoriasis diagnosis process through deep convolutional neural networks and an application for mobile devices. For classification, a dataset was created composed of 752 images of human skin, fed to 3 classification networks: MobileNetV2, Xception and InceptionResnetV2. A semantic segmentation step through 3 different architectures based on the topology of the U-net network was developed to assess the influence of segmentation on the classification. The results present a good potential of the system for the classification of psoriasis images. A set of tests between different types of datasets and neural networks was applied and pointed out that, although the proposed segmentation model reached more than 97% accuracy and a DICE score greater than 91%, providing a relatively good segmentation, the original dataset provides greater classification accuracy compared to the segmented dataset. In the classification, the MobileNetV2 model presented an accuracy greater than 97% with a loss rate of approximately 7%. The developed mobile application is able to use the classification and segmentation models in a simple way, assisting the end user in the process of analyzing the dermatological lesion through an intuitive interface.
Description
Mestrado de dupla diplomação com a UTFPR - Universidade Tecnológica Federal do Paraná
Keywords
Redes neurais Classificação Segmentação semântica Aplicação móvel