Diabetic retinopathy detection based on deep learning

• Ano de defesa: 2019
• Autor(a) principal: Zago, Gabriel Tozatto
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Engenharia Elétrica; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Imagens de retina; Aprendizado profundo; Retinopatia diabética; Redes neurais convolucionais; Qualidade de imagem; Localização de lesão; Retinal images; Deep learning; Diabetic retinopathy; Convolutional neural networks; Image quality; Lesion localization; subject.br-rjbn; Engenharia Elétrica
• Link de acesso: http://repositorio.ufes.br/handle/10/13661

Detecting the early signs of diabetic retinopathy (DR) is essential, as timely treatment might reduce or even prevent vision loss. Moreover, automatically localizing the regions of the retinal image that might contain lesions can favorably assist specialists in the task of detection. At the same time, poor-quality retinal images do not allow an accurate medical diagnosis, and it is inconvenient for a patient to return to a medical center to repeat the fundus photography exam. In this thesis, we argue that it is possible to propose a pipeline based on quality assessment and red lesion localization to achieve automatic DR detection with performance similar to experts considering that a rough segmentation is sufficient to produce a discriminant marker of a lesion. A robust automatic system is proposed to assess the quality of retinal images aiming at assisting health care professionals during a fundus photograph exam. We propose a convolutional neural network (CNN) pretrained on non-medical images for extracting general image features. The weights of the CNN are further adjusted via a fine-tuning procedure, resulting in a performant classifier using only with a small quantity of labeled images. We also designed a lesion localization model using a deep network patch-based approach. Our goal was to reduce the complexity of the implementation while improving its performance. For this purpose, we designed an efficient procedure (including two convolutional neural network models) for selecting the training patches, such that the challenging examples would be given special attention during the training process. Using the labeling of the region, a DR decision can be given to the initial image, without the need for special training. Our patch based approach allows the model to be trained with only 28 images achieving similar results to works that used over a million of labeled images. The CNN performance for quality assessment was evaluated on two publicly available databases (i.e., DRIMDB and ELSA-Brasil) using two different procedures: intra-database and inter-database cross-validation. The CNN achieves an area under the receiver operating characteristic curve (AUC) of 99.98% on DRIMDB and an AUC of 98.56% on ELSA-Brasil in the inter-database experiment, where training and testing were not performed on the same database. These results suggest the robustness of the proposed model to various image acquisitions without requiring special adaptation, thus making it a good candidate for use in operational clinical scenarios. The lesion localization model was trained on the Standard Diabetic Retinopathy Database, Calibration Level 1 (DIARETDB1) database and was tested on several databases (including Messidor) without any further adaptation. It reaches an area under the receiver operating characteristic curve of 0.912 - 95%CI 0.897-0.928 for DR screening, and a sensitivity of 0.940-95%CI 0.921-0.959. These values are competitive with other state-of-the-art approaches. The results suggest that the given hypothesis is confirmed.