Understanding inherent image features in CNN-based assessment of diabetic retinopathy

Abstract

Diabetic retinopathy (DR) is a leading cause of blindness and affects millions of people throughout the world. Early detection and timely checkups are key to reduce the risk of blindness. Automated grading of DR is a cost-effective way to ensure early detection and timely checkups. Deep learning or more specifically convolutional neural network (CNN)—based methods produce state-of-the-art performance in DR detection. Whilst CNN based methods have been proposed, no comparisons have been done between the extracted image features and their clinical relevance. Here we first adopt a CNN visualization strategy to discover the inherent image features involved in the CNN’s decision-making process. Then, we critically analyze those features with respect to commonly known pathologies namely microaneurysms, hemorrhages and exudates, and other ocular components. We also critically analyze different CNNs by considering what image features they pick up during learning to predict and justify their clinical relevance. The experiments are executed on publicly available fundus datasets (EyePACS and DIARETDB1) achieving an accuracy of 89 ~ 95% with AUC, sensitivity and specificity of respectively 95 ~ 98%, 74 ~ 86%, and 93 ~ 97%, for disease level grading of DR. Whilst different CNNs produce consistent classification results, the rate of picked-up image features disagreement between models could be as high as 70%.

Deep learning in the grading of diabetic retinopathy: A review

“…The authors of Ref. [1,41] divided the dataset into training, validation, and testing with the ratio of 80:10:10 and 94:3:3, respectively. In an automated system, it is important to remember that bias and prejudice must be avoided.…”

“…The authors of Ref. [1,25,31,42], and [29] perform transfer learning on multiple networks and compare the performance of each model. Dai et al [27] performs transfer learning and create DeepDR by combining three sub-networks that serve different purposes.…”

“…The authors of Ref. [1,24,25,31,36,40,42], and [29] employ different version of Inception architecture. Khalifa et al [42] achieved the highest accuracy 96.30% while the authors of Ref.…”

A Transfer Learning Approach for Diabetic Retinopathy and Diabetic Macular Edema Severity Grading

“…They use Messidor and IDRiD dataset for evaluation. Reguant et al [15] proposed a method where they initially perform visualization using CNN to find the innate image features engaged with the CNN's accountability process. Then, they fundamentally break down those provisions regarding generally known pathologies in particular hemorrhages, microaneurysms and exudates, and other visual segments.…”

A new detection model of microaneurysms based on improved FC-DenseNet

“…Kou et al proposed an architecture for U-Net obtained by combining the deep residual model and recurrent convolutional operations into U-Net 12 . Reguant et al proposed an unsupervised method for DR detection based on CNN 13 . González-Gonzalo et al proposed a deep visualization method based on the unsupervised selective inpainting 14 .…”