Abstract
Purpose :
To propose a robust method for the automatic grading of the DR using the ResNet-50 network with a modified layer architecture while emphasizing the importance of providing visual explanations.
Methods :
We used the more two largest publicy data sets. The EyePACS data set and te DDR data set. .EyePACS contains 88,702 fundus images but only the training set of images (35,126 images) was used, since only the labels of this group are publicly available. The DDR data set contains 13,673 images collected across 147 hospitals in China but after removing those of bad quality we only used 12,522 images from the DDR database.
To improve the method of authomatic grading we included additional deep learning techniques such as data augmentation, regularization, early stopping criteria, transfer learning and fine tuning and in order to assist in the interpretation of the results of the deep learning model, we introduced a visual Explainable Artificial Intelligence approach employing SHapley Additive exPlanations (SHAP).
Results :
The use of SHAP provided invaluable insights into the interpretability of the results. We achieved remarkable accuracy rates of 86.36% for EyePACS and 84.23% for DDR, reaffirming the efficacy of our approach in DR grading while simultaneously offering visual explanations through SHAP
Conclusions :
Our study demonstrates the effectiveness of the proposed method in accurately classifying the stages of DR from retinal fundus images. Moreover, this work overcomes the challenges of a highly imbalanced dataset, commonly encountered in clinical environments. Additionally,this research stands out, to the best of our knowledge, as the first instance of using SHAP for visual explanation in DR grading, which adds a significant novel contribution to the field.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.