Loading...
Publication
ClassificaĆ§Ć£o do glaucoma usando deep learning e imagens de fundo de olho obtidas com smartphone
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 39270 | 23.6 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Nos Ćŗltimos anos, os algoritmos de deep learning mostraram um potencial sem precedentes para auxiliar no diagnĆ³stico da doenƧa ocular glaucoma, considerada uma das principais causas de cegueira irreversĆvel no mundo. Esses algoritmos sĆ£o comumente usados em estudos para identificar essa doenƧa a partir de imagens digitais do fundo de olho, amplamente empregadas na maioria das clĆnicas oftalmolĆ³gicas e adquiridas por meio do exame tradicional de fundoscopia. AlĆ©m dos algoritmos de deep learning, o desenvolvimento tecnolĆ³gico permitiu a criaĆ§Ć£o de dispositivos portĆ”teis com capacidade de aplicar o exame fundoscĆ³pico por meio de smartphones, tornando o teste ainda mais acessĆvel. Diante dessa situaĆ§Ć£o, este trabalho coletou um novo conjunto de dados chamado Brasil Glaucoma (BrG), contendo 2.000 imagens da retina obtidas com o oftalmoscĆ³pio panoptic Wech Allyn, um dispositivo considerado simples e portĆ”til, com conexĆ£o de um adaptador para smartphone. Em seguida, treinou-se um algoritmo desenvolvido por meio de um ensemble de redes neurais convolucionais, de modo que a sua precisĆ£o na classificaĆ§Ć£o das imagens BrG em glaucomatosas e normais foi estimada. Assim, alĆ©m de um teste em imagens globais, ou seja, apresentando toda a Ć”rea fotografada, tambĆ©m foram realizados testes em trĆŖs tipos de regiƵes de interesse: a regiĆ£o do disco Ć³ptico, obtida por meio de segmentaĆ§Ć£o manual; a regiĆ£o especĆfica do disco Ć³ptico; e o disco Ć³ptico e Ć”reas adjacentes. Estes dois Ćŗltimos procedimentos foram aplicados de forma automatizada com auxĆlio de uma rede neural U-Net. Considerando a classificaĆ§Ć£o do glaucoma, o ensemble alcanƧou, no seu melhor desempenho, 95,4% de acurĆ”cia, 98,5% de sensibilidade e 98,2% de especificidade usando a regiĆ£o do disco Ć³ptico e adjacĆŖncias como entrada. Deste modo, apesar de limitaƧƵes como a falta de biomarcadores robustos e computĆ”veis para o diagnĆ³stico do glaucoma, concluiu-se que o algoritmo de deep learning desenvolvido possui capacidade substancial para a classificaĆ§Ć£o desta neuropatia Ć³ptica a partir do reconhecimento de padrƵes em imagens fundoscĆ³picas de baixa resoluĆ§Ć£o adquiridas atravĆ©s de um smartphone. Portanto, essas tecnologias de aquisiĆ§Ć£o de fotografias da retina combinadas com algoritmos de inteligĆŖncia artificial podem ter um grande impacto em pesquisas sobre o glaucoma e no diagnĆ³stico automatizado, contribuindo para a viabilidade de futuros programas de triagem populacional para essa doenƧa ocular.
In recent years, deep learning algorithms have shown unprecedented potential to help diagnose glaucoma eye disease, considered one of the main causes of irreversible blindness in the world. These algorithms are commonly used in studies to identify this disease from digital images of the fundus, widely used in most ophthalmological clinics and acquired through the traditional fundoscopy examination. In addition to deep learning algorithms, technological development has allowed the creation of portable devices that can apply the fundoscopic examination using a smartphone, making the test even more accessible. Given this situation, this work established a new dataset called Brazil Glaucoma (BrG), containing retinal images 2.000 obtained with the Welch Allyn panoptic ophthalmoscope, a device considered simple and portable with a smartphone adapter connection. Then, an ensemble of convolutional neural networks was developed and its accuracy in classifying BrG images into glaucomatous and normal was estimated. Thus, besides a test on global images, that is, showing the entire photographed area, tests were also performed on three types of regions of interest: the optical disc region, obtained through manual segmentation; the specific optical disc region; and the optical disc and adjacent areas. The last two procedures were applied in an automated way with the aid of a U-Net neural network. Considering glaucoma classification, the it ensemble achieved its best performance, 95.4% accuracy, 98.5% sensitivity, and 98.2% specificity using the optic disc region and adjacencies as input. Thus, despite limitations such as the lack of robust and computable biomarkers for the diagnosis of glaucoma, it was concluded that the developed deep learning algorithm has substantial capacity for the classification of this optic neuropathy from the recognition of patterns in low-resolution images acquired through a smartphone. Therefore, these fundoscopic image acquisition technologies combined with artificial intelligence algorithms can have a great impact on glaucoma research and automated diagnosis, contributing to the feasibility of future population screening programs for this eye disease.
In recent years, deep learning algorithms have shown unprecedented potential to help diagnose glaucoma eye disease, considered one of the main causes of irreversible blindness in the world. These algorithms are commonly used in studies to identify this disease from digital images of the fundus, widely used in most ophthalmological clinics and acquired through the traditional fundoscopy examination. In addition to deep learning algorithms, technological development has allowed the creation of portable devices that can apply the fundoscopic examination using a smartphone, making the test even more accessible. Given this situation, this work established a new dataset called Brazil Glaucoma (BrG), containing retinal images 2.000 obtained with the Welch Allyn panoptic ophthalmoscope, a device considered simple and portable with a smartphone adapter connection. Then, an ensemble of convolutional neural networks was developed and its accuracy in classifying BrG images into glaucomatous and normal was estimated. Thus, besides a test on global images, that is, showing the entire photographed area, tests were also performed on three types of regions of interest: the optical disc region, obtained through manual segmentation; the specific optical disc region; and the optical disc and adjacent areas. The last two procedures were applied in an automated way with the aid of a U-Net neural network. Considering glaucoma classification, the it ensemble achieved its best performance, 95.4% accuracy, 98.5% sensitivity, and 98.2% specificity using the optic disc region and adjacencies as input. Thus, despite limitations such as the lack of robust and computable biomarkers for the diagnosis of glaucoma, it was concluded that the developed deep learning algorithm has substantial capacity for the classification of this optic neuropathy from the recognition of patterns in low-resolution images acquired through a smartphone. Therefore, these fundoscopic image acquisition technologies combined with artificial intelligence algorithms can have a great impact on glaucoma research and automated diagnosis, contributing to the feasibility of future population screening programs for this eye disease.