{05/09} · Imaging

All work

CKD Detection from Fundus Photographs with RETFound

Role
Second author
Venue
PLOS Digital Health
Status
Under revision

Can a photograph of the retina flag chronic kidney disease?

We fine tuned RETFound on 1,472 patients imaged with a handheld camera. Grad-CAM, GradCAM++ and Integrated Gradients were compared by three clinical graders, one engineer and an automated script.

What the record supports

Claim

The model separates CKD from non CKD at the same hospital.

Evidence

Mean five fold patient level AUC was 0.825, and 0.937 on a second camera at the same hospital after colour harmonisation. Platt scaling cut external calibration error from 0.311 to 0.035.

Limits, stated plainly

AUC fell to 0.489 on two unrelated public datasets, and the paper reports this plainly.

Next project · warp onPreprocessing and Augmentation for CAD Detection