Fig. 3: Heatmaps generated for compressed images.

Heatmaps showing the ‘hot’ areas that the DL algorithm focuses its attention on when making a diagnostic assessment on the retinal image. This was created using the Integrated Gradient method⁶⁶. The colors on the greyscale retina image show the region of interest, with the red showing peak areas of region of interest while the blue shows the background areas of the region of interest. The white box isolates an area of the image to illustrate the difference between images of 350 and 150 KB in size. a A fundus photo of a healthy retina that was provided to the DL model as a 350 KB image. This was correctly classified by the DL model as a healthy retina with no DR. The heatmaps show focus on the normal optic disc and vasculature. b The same healthy retina is shown but compressed into a 150 KB size. This was misclassified by the DL algorithm as a retina with referable DR. The heatmaps show other regions of interest aside from the normal optic disc. The magnification of one of these anomalous regions of interest depicts pixelations as identified by the white arrows and ovals. These pixelations amalgamate into a mistaken pathological manifestation of DR, resulting in its false positive status.