Image quality and artifact feedback for fundus imaging using edge device for teleretinal application

Abstract

Purpose : In teleretinal settings, the instruments are operated by technicians who might not have the required expertise to decide whether the image is of sufficient quality for clinical review. Deep neural networks (DNN) have been shown to provide good performance for differentiating good from poor quality images. When the image is flagged as poor quality, it often does not give additional information about the types of artifact. In this project we developed a DNN to classify good/poor images and to give feedback on artifacts.

Methods : This research used retrospective study data of 5059 images acquired using VELARA™ 200 (ZEISS, Dublin, CA) camera. The images were acquired from normal and diseased subjects with various retinal pathologies. The images were graded for image quality (IQ) and presence of artifacts by a minimum of 2 graders and adjudicated by a subject expert. The images were graded for IQ in a 1-5 scale and converted to a binary scale for ground truths (GT) (0:IQ<=2, 1:IQ>2). The following artifacts were also marked in the images: small pupil artifact (SPA) and incorrect fixation artifact (IFA). The development set was split into training (3993 images with 69% good IQ, 16% SPA, and 19% IFA), testing (717 images with 66% good IQ, 16% SPA and 3% IFA) and validation (409 images with 59% good IQ, 17% SPA and 6% IFA).
SqueezeNet is a small DNN that is well suited for low-resource environments lacking GPUs. DNN is modified by adding a residual layer, input channels: RGB image, Hessian determinant, Ridge map and black channel and three-class outputs: IQ, SPA and IFA. The network is trained using weighted cross-entropy to account for class imbalance. The outputs from the DNN are compared with the GTs in test set and performance metrics such as ROC, precision-recall curves, accuracy, area under the curve (AUC) are reported.

Results : Figure 1 shows some examples of good and poor quality images. Figure 2 shows the ROC curves for each of the outputs. The proposed DNN achieved accuracies of 91.9%, 94.1% and 94.4% and AUCs of 0.97, 0.98 and 0.94 for IQ, SPA and IFA respectively.

Conclusions : To help provide useful feedback to the untrained operator, we developed a three-class classification using modified SqueezeNet. The results show potential to serve as image quality and artifact detector for teleretinal applications. Future studies will involve collecting more images to fine-tune the performance.

This is a 2021 Imaging in the Eye Conference abstract.