Diabetic Retinopathy Detection Based On Deep Learning
Nome: GABRIEL TOZATTO ZAGO
Tipo: Tese de doutorado
Data de publicação: 20/12/2019
Orientador:
Nome |
Papel |
|---|---|
| EVANDRO OTTONI TEATINI SALLES | Orientador |
Banca:
| Nome |
Papel |
|---|---|
| AURA CONCI | Examinador Externo |
| EVANDRO OTTONI TEATINI SALLES | Coorientador |
| MARIANA RAMPINELLI FERNANDES | Examinador Externo |
| PATRICK MARQUES CIARELLI | Examinador Externo |
| RODRIGO VAREJÃO ANDREÃO | Orientador |
Páginas
Resumo: Finding early signs of DR is essential because the early treatment might reduce or even stop the patients vision loss. Localizing the regions of the retinal image that might contain lesions can assist the specialists in reducing the burden of their work or even increasing their hit rate. Deep learning models have become state of the art in many computer vision tasks and it one of its advantages is to reduce the empirically chosen parameters such as filters
parameters, thresholds and others which might lead the method to be database dependent. In this work, a patchbased approach has been described including two CNN models - one to select patches in which labeling a lesion may be ambiguous and a second one, trained with the previously selected patches, to accomplish the final classification. The model is trained on DIARETDB1 database and tested on MESSIDOR without any further adaptation reaching area under the ROC curve of 0.912 ± 0.016 for diabetic retinopathy screening.
