Abstract
Purpose :
Diabetic retinopathy (DR) is a leading cause of blindness worldwide. Early diagnosis and reliable detection of progression are critical for better prognosis. Currently, DR is classified by severities but evaluating clinical features directly may improve the accuracy of classifications and provide important insights in DR pathology or progression. We developed a deep learning pipeline that not only classifies DR severity, but also quantifies all major DR features.
Methods :
55 fundus images were manually segmented by reading center experts and used to train a deep segmentation network that locates all major diabetic retinopathy features, namely microaneurysms, hemorrhages, exudates, cottonwool spots, neovascularization, intraretinal microvascular abnormality and venous beading, at high resolution. To investigate the usefulness of the segmentations, this network was used to segment DR features in Kaggle Diabetic Retinopathy competition (EyePACS). We then trained three different classification networks: a traditional convolutional neural network, and two pyramidal networks with fractional max pooling (a shallower and a deeper model), to predict DR severity on the Kaggle data with and without the segmentation masks.
Results :
Representative segmentations from the validation set are shown in Figure 2. Higher classification accuracy were achieved with masks (69.2%, 66.2%, 70.1%) than without (66.9%, 63.2%, 68.9%) for the 3 networks. Random forest ensembling improved accuracies to 83% vs 78% for with and without respectively.
Conclusions :
We present a new AI pipeline that improves the accuracy of DR classification by incorporating DR feature segmentation. The pipeline yields segmentations that can help clinicians identify and locate clinically relevant DR features, thereby enabling a new generation of feature-based studies of DR diagnosis and progression rather than the traditional severity scheme, leading to novel hypotheses and analyses.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.