Purchase this article with an account.
Kaushal Solanki, Sandeep Bhat, Chaithanya Ramachandra, Muneeswar Gupta Nittala, Srinivas R Sadda; EyeArt: Advanced Image Analysis Tools for Diabetic Retinopathy Screening and Telemedicine Applications. Invest. Ophthalmol. Vis. Sci. 2014;55(13):5883. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Developing and deploying computerized screening algorithms is the only way to screen large, growing diabetic population for preventable blindness due to diabetic retinopathy (DR). To achieve this, we have developed EyeArt, a computerized DR screening system with high diagnostic efficacy, and high generalizability demonstrated via conducting cross-dataset testing. System design includes potential telemedicine integration via web-based application programming interface (API).
Our novel retinal color fundus image analysis framework comprises of the following steps: (i) image enhancement that normalizes intensity and other variations, (ii) interest region detection that selects < 1% of pixels for further analysis, (iii) descriptors set computation in a multi-scale framework, (iv) support vector machine classification at pixel, lesions, and image level. The resultant end-to-end DR screening system is engineered to work on the cloud cluster (Amazon Elastic Cloud). We evaluate EyeArt on a large 1200-image public dataset (Messidor) for two different scenarios: (a) detecting any signs of DR, and (b) detecting DR onset as defined by >5 microaneurysms (MA) or >0 hemorrhages (HM). Both “any DR” and “DR onset” were tested in two scenarios. First, we test a 50-50 split of the Messidor dataset allowing comparison with many existing approaches, and second, we conduct cross-dataset test where all 1200 images are tested upon while training on a completely different, much smaller, USC-DEI dataset.
We achieve AUROC of 0.91 for “any DR” and 0.95 for “DR onset” as shown in Figure 1, which is significantly better than all competing approaches on the same Messidor dataset (Table 1). It also matches human graders’ performance on the same dataset. EyeArt also performs reasonably well (AUROC of 0.86 for “any DR” and 0.91 for “DR onset”) for cross-dataset testing.
We present EyeArt, a new approach for DR screening that has potential for real-life use with its demonstrated high diagnostic efficacy (improving upon the current state of the art and matching human graders' performance), good generalizability (good cross dataset testing results), explicit invariance to imaging conditions (via illumination normalization), and flexible packaging (web-based API and linkable libraries for telemedicine integration).
This PDF is available to Subscribers Only