Purchase this article with an account.
Paras Vora, Eric Higgins; Inexpensive Deep Learning for Semantic Vessel and Lesion Segmentation in Diabetic Retinopathy. Invest. Ophthalmol. Vis. Sci. 2020;61(7):2013.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Diabetic retinopathy is an increasingly prevalent disease, difficult to screen for across the globe. Clinicians currently screen for this disease by correlating vascular findings, as well as the size, presence and location of lesions such as hard exudates and microaneurysms. Automatic segmentation of the vasculature and lesions found in fundus images is a promising avenue for accurate, early, and efficient clinical detection. Deep learning methods to date have not found their place in clinical practice. Additionally, the training of these models has traditionally been cost-prohibitive. We propose a novel application of a neural network design, trained on the free Google Colaboratory (Colab) platform, to perform low-cost semantic segmentation of fundus photos that extracts vasculature and lesions to enhance clinician attention in diabetic retinopathy screening.
We developed a fully convolutional neural network based on the U-Net design. Retinal images (40-50 per model) from databases were used to train the model for vasculature and lesion (microaneurysm, hemorrhage, hard exudate, soft exudate) segmentation. These databases contained pixel-level annotations. These models were trained exclusively using Google Colab, leveraging free high-performance computation. Each model was trained for 85 epochs, using categorical cross-entropy loss. The performance of the algorithms was evaluated by comparing results with images not used during training or validation.
The Area Under Curve (AUC) of the Receiver Operating Characteristic (ROC) curve for the vasculature, hard exudate, microaneurysm, hemorrhage, and soft exudate datasets was 0.9789, 0.9843, 0.9415, 0.9523 and 0.9761 respectively. The accuracy of each model for vasculature, hard exudate, microaneurysm, hemorrhage, and soft exudate datasets was 95%, 97%, 95%, 85% and 94% respectively.
The proposed models provide promising performance using relatively little training data for semantic segmentation of vasculature and lesions in fundus photos of patients with diabetic retinopathy. The proposed approach demonstrates that state-of-the-art results are readily accessible using free platforms. With more data and a more optimized model architecture, we would expect more performant results. We hope to apply these models in an ensemble fashion to see if they aid clinicians in early detection of diabetic retinopathy.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only