Purchase this article with an account.
Michelle Y.T. Yip, Gilbert Lim, Valentina Bellemo, Yuchen Xie, Xin Qi Lee, Quang Nguyen, Haslina Hamzah, Jinyi Ho, Charumathi Sabanayagam, Carol Yim-lui Cheung, Gavin Siew Wei Tan, Wynne Hsu, Mong Li Lee, Tien Yin Wong, Daniel SW Ting; Effect of Image Compression and Number of Fields on a Deep Learning System for Detection of Diabetic Retinopathy. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1438.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Many deep learning systems(DLS), trained with different methods, reported excellent diagnostic ability to detect diabetic retinopathy(DR). As different screening programs around the world produce images of varying specifications such as image size and field of view, it is important to evaluate the generalizability of a DLS to accommodate for these variations. This study aims to investigate the impact of the image compression and field of view on the diagnostic performance of a previously validated DLS.
The DLS was trained with 2-field(45° macula(MC) and optic-disc centred(OD)) retinal images from the Singapore Integrated Diabetic Retinopathy Program(SiDRP) taken from 2010-13. It was trained to detect referable DR, defined as moderate non-proliferative DR or worse, diabetic macular edema or ungradable. To examine image compression, we resized raw retinal images collected from SiDRP between 2014-15, from 350 kilobytes(KB) to 300, 250, 200 and 150 KB using OpenCV’s JPEG compression engine(n=71,896 images, 35,948 eyes, 14,880 subjects). To explore fields of view, we evaluated the DLS’ performance on 1-field(45° MC) images from SiDRP 2014-15 compared to 2-field images. The outcome measures were area under the curve(AUC), sensitivity and specificity.
AUC of the DLS remained consistently above 0.9 despite reduction in image size and there is no significant difference until the image downsized to 250KB (p<0.05). Sensitivity and specificity is maintained above 80% until the threshold of 250KB (90.0% & 81.1% respectively) where further downsizing led to specificity to fall below 80%. Heatmap analysis showed that increase in false positives were due to lower image quality leading the DLS to focus on fake aberrancies. Providing the DLS with 1-field images preserves the performance above 0.85 (AUC 0.866, Sensitivity 91.7%, Specificity 91.9%) although a fall in AUC and specificity occurs (p<0.05).
Reducing image size and number of fields do not significantly affect the clinical diagnostic performance of the DLS, provided that image size is maintained to a minimum of 250KB and the single-field provided is a MC image.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Table 1. Performance of the DLS with different image sizes (top) and different fields of view (bottom).
Figure 1. Heatmap analysis of a normal retinal image misidentified to be diseased at a lower resolution due to the creation of fake aberrancies.
This PDF is available to Subscribers Only