Fig 1.
Representative fundus images with different pathological severity of DME.
(A) Denotes the images of Grade 0 severity which have no hard exudates, (B) Represents Grade 1 severity which has exudates outside the radius of one disc diameter from the macula center, and (C) Denotes the fundus under the Grade 2 severity category, with exudates within the radius of one disc diameter from the macula center.
Table 1.
Description of DME severity grading.
Fig 2.
Reprenstative fundus images depicting our augmentation techniques.
The images marked under column (A) are the original fundus images that are to be augmented. The fundus images under columns (B), (C) are the images generated owing to flipping and random rotations ranging from 10° to 40° respectively.
Fig 3.
Representative fundus images showing a clear distinction between original and preprocessed images.
The images in the column (A) are the raw fundus images and the images in the column (B) show the preprocessed images wherein we can observe the exudates (white/pale yellow spots) and blood vessels more clearly with the controlled background noise.
Fig 4.
Illustration of the workflow of proposed model—DMENet.
DMENet is a two-stage pipeline, where, in the first stage, the algorithm detects the presence/absence of DME and once the presence of DME is confirmed it is passed through the second stage where the image is graded based on the severity.
Fig 5.
Illustration of the proposed HE-CNN model (Two-level architecture).
Table 2.
CNN structures.
Fig 6.
Confusion Matrices showing the performances of (A) Gamma and (B) Delta Ensemble.
Fig 7.
The Receiver Operating Characteristic (ROC) curves demonstrating the performance of Gamma and Delta Ensemble.
The area under the curve is 0.9654 in the Gamma Ensemble and 0.9489 in the Delta ensemble.
Table 3.
Performance of Gamma and Delta Ensembles.
Table 4.
Comparitive study of HE-CNN ensemble.
Table 5.
Comparitive study of DMENet’s performance with recent solutions for DME screening using MESSIDOR dataset.
Table 6.
Comparitive study of DMENet’s performance with recent solutions for DME screening using IDRiD dataset.
Table 7.
Results of optimization methods.
Table 8.
Results of DMENet v/s tri-class classification.