Dawoon Leem, Hyungyu Lee, Synho Do, and Sangchul Yoon


ARVO 2020 / 2020.06


To ensure the reliability of the diagnosis for fundus image, some researches were proposed which classify the image quality as good and bad. However, there are many images that can be diagnosed among the images classified as bad quality. We propose a method that assesses the image quality score in fundus images with a three-grade quality (Good, Usable and Reject) and filters out the images with very poor quality classified as the unreadable grade.



In experiments, we used 28,792 fundus images of the EyePACS dataset to train QA model. In order to classify the image quality into the three grades, we combine the features to measure the local quality and global brightness, and morphological disc information. The final quality score is estimated by our transformation equation that convert probabilities of three classes obtained by the combined features and XGBoost to normalized scalar value. To detect unreadable grade of fundus images, we measure q-th quantile of the data along the quality score and split diabetic retinopathy (DR) classification accuracy of images corresponding each quantile into two groups. Decision boundary of unreadable grade is the quantile boundary of two clusters with two connected components. We use the average of the quantile boundaries of multiple q-th.

Fig. 1. Examples of images classified as the unreadable grade.



Our QA method achieved F1-score 87.82, it is 2.31 higher than state-of-the-arts. The decision boundary of unreadable grade of the DR model was estimated as 0.000552, and we detected 253 images with unreadable grade in 16,249 test images. The DR accuracy of the data classified as the unreadable grade was 6.89% lower than the accuracy of the readable grade. As shown in Fig. 1, images in the unreadable grade were out of focus and had very low or high brightness and low contrast.



We found that the performance of the DR classifier decreased even though a deep learning-based approach if the image quality is very poor. Therefore, our QA outperformed in comparison to the state-of-the-art. Moreover, our unreadable grade classification method can assure image quality that does not affect the diagnosis. If the image quality in other modality such as X-ray can be measured, our unreadable grade classification can be used to detect images with very poor quality that affect diagnosis.