Image quality assessment of ultra-widefield fundus images using deep convolutional neural networks.

Abstract

Purpose : Insufficient image quality of fundus images results in impairment of manual or automatic image grading process, resulting in loss of a data for that particular visit. Automated evaluation of image quality would help the operator by giving a quick feedback to reacquire the images when image quality is inadequate. In this paper, we propose a deep neural network based approach that predicts the quality of a high resolution fundus image immediately after acquisition.

Methods : We trained a deep Residual Network (ResNet) using TensorFlow for the task of image quality prediction. For model training, we used 561 good & 74 bad quality images taken using CLARUS^TM 500 (Zeiss, Dublin, CA ). The test set has 270 images of good quality and 28 of inadequate quality. Bad quality images used for training is about ~12%, which is greater than the average occurrence of bad quality images in any random image set, to address the issue of unbalanced classes in training data. The most common artifacts which obscure the clinically useful information and interfere with the ability to interpret the image are severe blurring, reflex, vignetting, striping, or rainbow discoloration as shown in Fig. 1. Given the very small dataset and unbalanced sample size, we use a data augmentation step at the time of model training by applying random flips to the images. To speed up the training and inference, we downscaled the high-resolution images and cropped each image to 224 x 224 pixel each.

Results : The results are shown in Fig. 2. The algorithm achieves an Area under the Curve (AUC) of 97.20%. Out of the 9 misclassified images, 66% of them are bad quality images, which are classified as good quality images.

Conclusions : We present a solution for automated image quality assessment of fundus images taken by CLARUS 500. The presented deep learning algorithm achieves a very promising result even with a handful of data. Thus, it enables the operator to obtain a feedback on image quality at acquisition time and to reacquire images when image quality is poor.

This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.

 

Fig. 1: Shows images used in model training (a) Good Quality Images (b) Bad Quality Images

 

Fig. 2: Results for Automated Fundus Image Quality Assessment.