Purchase this article with an account.
Mhd Hasan Sarhan, Katherine Makedonsky, Meike Mack, Mary Durbin, Mehmet Yigitsoy, Abouzar Eslami; Deep learning for automatic diabetic retinopathy detection under multiple image quality levels. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB0105.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Deep learning techniques are showing promising results for automatic diabetic retinopathy (DR) screening on fundus images. We analyze the effect of image quality of images collected with a handheld fundus camera on the performance of a DR screening algorithm.
This retrospective study used 257 fovea-centered fundus images from 93 subjects collected using VISUSCOUT® 100 (ZEISS, Jena, Germany) handheld fundus camera. The images are annotated positive for DR if signs of DR are visible (mild or more in International Clinical Diabetic Retinopathy disease severity scale). Image quality is assessed using a subjective 1-5 scale (1-very poor; 2-poor; 3-fair; 4-good; 5-excellent). Fig 1 shows an example of each quality level. 233 images were annotated as negative for DR (healthy) and 24 were annotated as positive for DR. We used our Diabetic Retinopathy Deep Network grading model (DRDN) that is trained on ~35k publicly available fundus images for grading the severity of DR. DRDN is used to classify VISUSCOUT images for Healthy vs DR by regarding a mild or more prediction as DR. No data from the same handheld device has been used for training the model. Sensitivity (Sn), specificity (Sp) and area under the curve of receiver operating characteristics curve (AUC) were reported for all images. Sn and Sp were used to analyze the effect of image quality on DR prediction performance.
DRDN shows good prediction power for unseen images from handheld camera reaching AUC of 0.98 on images from all quality profiles. No false negatives were triggered, hence, Sn=1. The model gave 22 false positives with Sp=0.905, 95% Confidence Interval (CI) [0.866, 0.943]. The results are shown in Fig 2.
A high number of false positives (Sp=0.8, 95% CI [0.656, 0.943]) occurred in very poor quality Images. For images of poor and above quality, we observe a consistent performance of the algorithm with less effect of image quality on the Sp. Using samples from fair quality and above gives the highest results with Sp=0.926, 95% CI [0.886, 0.966] which suggests that using images of above-poor quality would decrease the number of false positives in the DR screening.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.
Examples of fundus images for each quality level
Sn and Sp of 5 quality levels and all levels combined; vertical black lines show 95% CI; the amount of available images for each quality is in parentheses; red dotted horizontal line is Sp over all images
This PDF is available to Subscribers Only