Deep learning for automatic diabetic retinopathy grading of ultra-widefield fundus images

Abstract

Purpose : Ultra-widefield (UWF) fundus photography provides a powerful tool for monitoring the periphery of the retina compared to narrow field (NF) images. We study the effectiveness of transfer learning based deep learning techniques for automatic diabetic retinopathy (DR) grading in ultra-widefield fundus images using a small amount of annotated UWF images.

Methods : We collected 154 UWF fundus images using CLARUS^TM 500 (ZEISS, Dublin, CA). All images were graded for DR by a trained ophthalmologist using International Clinical Diabetic Retinopathy (ICDR) disease severity scale. The data consisted of a mix of healthy and different DR levels images captured from 119 subjects. We used an inception-resnet neural network to create an automatic DR grading model using NF kaggle fundus images (35k images, 45^ο FOV). We fine-tuned the NF model by using the UWF images. We used various data augmentation techniques to account for data scarcity. The model was evaluated using 5-fold stratified cross-validation. We used accuracy and quadratic weighted kappa score for grading evaluation. Kappa, accuracy, sensitivity, specificity, and area under the curve of receiver operating characteristics curve (AUC) were used for screening (Healthy vs DR) and referable DR detection (More than mild DR) evaluation. Each UWF image is divided into 4 non-overlapping tiles and the results are combined using the mean prediction value. We compare the usage of CLARUS data for fine-tuning the model with using the pre-trained model on NF data for evaluation.

Results : The results are shown in Table 1. The model reached an accuracy of 84%, 97% and 94% for DR grading, referable DR, and DR screening respectively. This is an improvement over using only NF images in all considered metrics and it is achievable using a small amount of UFW images.

Conclusions : The current scarcity of annotated UWF fundus images can be addressed using transfer training given the large number of available annotated NF fundus images and fine tuning with a small amount of UWF images. This approach shows a good generalization of an existing trained model without the need for large-scale data collection. In the future, the performance of the algorithm is expected to improve as more annotated UWF images are added to the training set.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Table 1 Comparison between using CLARUS data for fine-tuning and using CLARUS data directly for evaluation on the NF trained model