Loading...
Publication
Doenças hereditárias da retina: aumento de dados, utilizando métodos generativos, para melhorar a deteção
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 1.61 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
As doenças hereditárias da retina (IRDs) constituem um grupo de patologias raras e geneticamente heterogéneas, responsáveis pela perda progressiva da visão. A identificação do gene causador é essencial para o tratamento, mas requer testes genéticos extensos e dispendiosos. Neste contexto, a aplicação de técnicas de deep learning a exames de Autofluorescência de Fundo Ocular (FAF) surge como uma alternativa promissora para auxiliar a classificação genética destas doenças.
Contudo, a raridade das IRDs traduz-se na falta de conjuntos de dados amplos e equilibrados, dificultando a representação adequada de todas as classes genéticas. Essa limitação pode afetar significativamente a capacidade dos modelos de deep learning aprenderem padrões representativos e generalizáveis.
Esta dissertação teve como objetivo avaliar o desempenho de modelos de deep learning na classificação de IRDs e explorar o potencial de métodos generativos para o aumento de dados em conjuntos de imagens reduzidos e desbalanceados.
Foram comparados três modelos do estado da arte (VGG19, Inception-V3 e ResNet-50) utilizando validação cruzada (5-fold). O modelo ResNet-50 apresentou o melhor desempenho, com acurácia de 56% e F1-score de 0,54, revelando boa capacidade discriminatória em classes com maior representatividade.
Numa segunda fase, foi realizada uma revisão da literatura com o objetivo de identificar os métodos generativos utilizados para aumento de dados nas IRDs, bem como uma implementação da arquitetura DCGAN para gerar imagens FAF sintéticas e mitigar o desbalanceamento entre classes. Apesar de as imagens geradas não terem atingido a qualidade visual desejada, a implementação comprovou a viabilidade da abordagem generativa. Além disso, a revisão evidenciou o sucesso destas técnicas em outras doenças da retina, onde foram reportados resultados muito favoráveis.
Inherited retinal diseases (IRDs) are a group of rare and genetically heterogeneous pathologies responsible for progressive vision loss. Identifying the causative gene is essential for treatment but requires extensive and costly genetic testing. In this context, the application of deep learning techniques to fundus autofluorescence (FAF) examinations emerge as a promising alternative to aid in the genetic classification of these diseases. However, the rarity of IRDs translates into a lack of large and balanced datasets, making it difficult to adequately represent all genetic classes. This limitation can significantly affect the ability of deep learning models to learn representative and generalizable patterns. This dissertation aimed to evaluate the performance of deep learning models in the classification of IRDs and explore the potential of generative methods for data augmentation in small and unbalanced image sets. Three state-of-the-art models (VGG19, Inception-V3, and ResNet-50) were compared using cross-validation (5-fold). The ResNet-50 model performed best, with an accuracy of 56% and an F1-score of 0.54, revealing good discriminatory ability in classes with greater representativeness. In a second phase, a literature review was conducted to identify the generative methods used for data augmentation in IRDs, as well as an implementation of the DCGAN architecture to generate synthetic FAF images and mitigate class imbalance. Although the generated images did not achieve the desired visual quality, the implementation proved the feasibility of the generative approach. In addition, the review highlighted the success of these techniques in other retinal diseases, where very favorable results have been reported.
Inherited retinal diseases (IRDs) are a group of rare and genetically heterogeneous pathologies responsible for progressive vision loss. Identifying the causative gene is essential for treatment but requires extensive and costly genetic testing. In this context, the application of deep learning techniques to fundus autofluorescence (FAF) examinations emerge as a promising alternative to aid in the genetic classification of these diseases. However, the rarity of IRDs translates into a lack of large and balanced datasets, making it difficult to adequately represent all genetic classes. This limitation can significantly affect the ability of deep learning models to learn representative and generalizable patterns. This dissertation aimed to evaluate the performance of deep learning models in the classification of IRDs and explore the potential of generative methods for data augmentation in small and unbalanced image sets. Three state-of-the-art models (VGG19, Inception-V3, and ResNet-50) were compared using cross-validation (5-fold). The ResNet-50 model performed best, with an accuracy of 56% and an F1-score of 0.54, revealing good discriminatory ability in classes with greater representativeness. In a second phase, a literature review was conducted to identify the generative methods used for data augmentation in IRDs, as well as an implementation of the DCGAN architecture to generate synthetic FAF images and mitigate class imbalance. Although the generated images did not achieve the desired visual quality, the implementation proved the feasibility of the generative approach. In addition, the review highlighted the success of these techniques in other retinal diseases, where very favorable results have been reported.
Description
Mestrado em Engenharia Informática e Tecnologia Web, apresentada à Universidade Aberta
Keywords
Modelos generativos Aumento de dados Dados sintéticos Doenças da retina Doenças hereditárias da retina Generative models Data augmentation Synthetic data Retinal diseases Inherited retinal diseases
Pedagogical Context
Citation
Publisher
CC License
Without CC licence