Abstract
Purpose :
Diabetic retinopathy (DR) is one of the most common causes of blindness within diabetic population. Screening using fundus images helps to detect DR at early stages. Various types of fundus cameras are available today, which exhibit stark technical differences, e.g., with respect to image resolution, colorization, illumination, and acquisition quality in general. To screen a large population, it is essential to develop an automated DR screening system that is not tailored to a specific fundus camera device but is instead reliably applicable to multiple device types.
Methods :
We compare two complementary approaches: learned preprocessing (LPP) and task-inspired preprocessing (TIPP). For LPP, images are resized to constant width and height and directly presented to a convolutional neural network (CNN). Thereby, the CNN needs to learn adequate preprocessing. As TIPP, we build on the well-known Graham algorithm and modify it such that normalization is applied after disc cropping to avoid device-dependent normalization inconsistencies. Resizing is done to a disc’s normalized radius of 125px. A radius of 15px is used for local mean subtraction as a compromise between speed and visual appearance.
We train InceptionResnet V2 from scratch using Tensorflow. We restrict training data to 133k images recorded with VISUSCOUT® 100 (ZEISS, Jena, Germany), a hand-held low-cost fundus device. As in-device-test, we evaluate trained models on a 15k VISUHEALTH® hold-out test set. As device-generalization-test, we evaluate trained models on E-Optha and Messidor-2 (open source fundus data sets) which were recorded with tabletop cameras.
Results :
Prediction accuracy for DR screening is shown in Fig 2. LPP drastically falls short for data from different devices (blue middle and right). In contrast, TIPP barely reduces accuracy on the original device type (orange left), but clearly lifts generalization performance (orange middle and right).
Conclusions :
From the perspective of general screening scenarios, we conclude that task-inspired preprocessing may be applied to allow for reliable cross-device predictions.
This is a 2020 Imaging in the Eye Conference abstract.