Abstract
Purpose :
Diabetic retinopathy (DR) is an eye disease that is retinal microvascular damage due to diabetes mellitus. DR can lead to severe vision loss for patients. In this study, we propose an attention model using the deep convolutional neural network (DCNN) for the prediction of the critical regions of DR, which could predict attention maps for subsequent grading.
Methods :
Figure 1 displays the DCNN structure. Inspired by Gradient-weighted Class Activation Mapping (Grad-CAM), we design a visual explanation of DCNN, which could highlight the important regions for classification model via guided backpropagation. In the premise of the image attention maps, an end-to-end regression model was used to learn attention maps. The performance of grading DR was improved by creating attention weight using a trained attention model, as well as “re-weight” the original image.
Results :
We utilize 35126 images from Kaggle DR Detection Dataset to train the DCNN model. To measure if the attention maps help ophthalmologist screening for DR, we verify our method on the IDRiD dataset. We utilize 35126 images from Kaggle DR Detection Dataset to train the DCNN model. To measure if the attention maps help ophthalmologist screening for DR, we verify our method on the IDRiD dataset. Quadratic weighted kappa (QWK) for DR grading of 0.75 and 0.79 for without attention maps and with attention maps, respectively.
Conclusions :
The attention map obtained by the classification model is a region that is meaningful for DR grading, which is consistent with our hypothesis that attention maps highlight the diseased area. The proposed attention model learns to generate reliable attention regions for clinical interpretation without known grading scales.
This is a 2020 ARVO Annual Meeting abstract.