Deep residual transfer learning for automatic diagnosis and grading of diabetic retinopathy

NEUROCOMPUTING

卷:452

Article

Martinez-Murcia, Francisco J.^1,3  Ortiz, Andres^1,3  Ramirez, Javier^2,3  Gorriz, Juan M.^2,3  Cruz, Ricardo¹ 

[1] Univ Malaga, Dept Commun Engn, Malaga, Spain

[2] Univ Granada, Dept Signal Theory Commun & Networking, Granada, Spain

[3] Andalusian Res Inst Data Sci & Computat Intellige, Granada, Spain

关键词: Deep learning;    Residual learning;    Transfer learning;    Convolutional neural network;    Retinography;    Diabetic retinopathy;   

DOI  :  10.1016/j.neucom.2020.04.148

来源: Elsevier

【 摘 要 】

Evaluation and diagnosis of retina pathology is usually made via the analysis of different image modal-ities that allow to explore its structure. The most popular retina image method is retinography, a tech-nique that displays the fundus of the eye, including the retina and other structures. Retinography is the most common imaging method to diagnose retina diseases such as Diabetic Retinopathy (DB) or Macular Edema (ME). However, retinography evaluation to score the image according to the disease grade presents difficulties due to differences in contrast, brightness and the presence of artifacts. Therefore, it is mainly done via manual analysis; a time consuming task that requires a trained clinician to examine and evaluate the images. In this paper, we present a computer aided diagnosis tool that takes advantage of the performance provided by deep learning architectures for image analysis. Our proposal is based on a deep residual convolutional neural network for extracting discriminatory features with no prior complex image transformations to enhance the image quality or to highlight specific structures. Moreover, we used the transfer learning paradigm to reuse layers from deep neural networks previously trained on the ImageNet dataset, under the hypothesis that first layers capture abstract features than can be reused for different problems. Experiments using different convolutional architectures have been car-ried out and their performance has been evaluated on the MESSIDOR database using cross-validation. Best results were found using a ResNet50-based architecture, showing an AUC of 0.93 for grades 0 + 1, AUC of 0.81 for grade 2 and AUC of 0.92 for grade 3 labelling, as well as AUCs higher than 0.97 when con -sidering a binary classification problem (grades 0 vs 3). (c) 2020 Published by Elsevier B.V.

【 授权许可】

Free