Retinal microvascular biomarker extraction on fundus images from the Maastricht study using supervised deep learning

Retinal fundus imaging enables detailed visualization of the microvascular structure in the retina of the human eye. Geometrical features, related to vessel caliber, tortuosity and bifurcations, have been identified as potential biomarkers for a variety of A.J, including (pre)diabetes and hypertension. A pipeline of automated unsupervised image analysis methods for extraction of such features from retinal fundus images has previously been developed and evaluated [1]. However, the current computationally expensive pipeline takes 24 minutes to process a single image, which impedes implementation in a screening setting. In the present work, we approximate the pipeline using a deep neural network that enables processing of a single image in a few seconds. We use a model that contains approximately 23 million trainable parameters and we train it with color fundus images from the Maastricht Study, a population-based cohort study with extensive phenotyping, that focuses on the etiology, complications and comorbidities of Type 2 Diabetes Mellitus. The set comprises 10668 images from 2872 subjects taken from both left and right eyes and are centered either on the fovea or on the optic disc.We design the model to simultaneously output four global biomarkers that represent key vessel geometries: Central Retinal Arteriolar Equivalent (CRAE), Central Retinal Venular Equivalent (CRVE), global tortuosity and asymmetry ratio of the bifurcations. The outputs from the original pipeline are used as training labels. Eighty percent of the data is used for training, while the remainder is used to evaluate the performance of the model.We obtain a substantial speed-up, requiring only 5 seconds to process an image. Intraclass correlation coefficient between the predictions of the model and the results of the pipeline showed strong correlation (0.86 - 0.91) for three of four biomarkers and moderate correlation (0.42) for one biomarker. To visualize what regions in the fundus images contribute to the model predictions, we create class activation maps. The maps show clearly that the local activations overlap with the vascular tree. It is able to differentiate between arterioles and venules around the optic disc when predicting CRAE and CRVE. Moreover, local high and low tortuous regions are clearly identified, verifying that the model is sensitive to key structures in the retina