Diabetic retinopathy detection in widefield images by using transfer learning from handheld narrow field images

Abstract

Purpose : Fundus widefield (WF) images are rich with information related to the peripheral retina compared to the narrow field (NF) fundus photography. NF fundus cameras have been in use for long; hence, there exist larger amounts of NF data compared to WF. In this study, we show how well a deep learning system (DLS) can generalize from NF lower quality fundus images to WF high quality images for the task of referable diabetic retinopathy (DR) detection.

Methods : For the first stage, we collected 32k NF images using VISUSCOUT® 100 (ZEISS, Jena, Germany) handheld fundus camera. The images were annotated for referable DR (more than mild DR). A DLS was developed to detect referable DR from the NF images. For the transfer learning, we collected WF fundus images from 361 subjects (23% with referable DR) using tabletop CLARUSTM 500 (ZEISS, Dublin, CA). Images were annotated for referable DR by at least two experts. We split the WF dataset into train/validation/test sets and evaluated the final model on the test set. We compared the performance of three models. 1) DLS (ResNet-50) trained only on WF images from scratch; 2) NF DLS model without re-training to evaluate the classification of WF images; 3)Transfer learning, the NF model is used as an initial model to train the WF train set. Multiple augmentation techniques were used to account for the small amount of data. We compared the accuracy, sensitivity, specificity, and area under the curve of the receiver operating characteristics curve (AUC) for evaluation of the performance.

Results : The results are shown in Figure 2. The model reached an AUC of 0.91 for referable DR detection compared to 0.42, and 0.61 for training from scratch and directly using NF DLS respectively. This shows improvement over not using NF images or using NF images only in all considered metrics.

Conclusions : It is advantageous to use lower quality NF fundus images as a pre-training step for a DLS to enhance the results of WF fundus image classification. This is important due to scarcity of WF images compared to NF images. The method shows good generalization from NF to WF without collecting a prohibitive amount of new data.

This is a 2020 Imaging in the Eye Conference abstract.

 

Figure 1. Transfer learning framework

 

Figure 2. Comparison between training from scratch, training only NF, and transfer learning from NF to WF