A Deep Learning Based Architecture to Diagnose Diabetes Using Retinal Images

Abstract

Diabetes is one of the leading fatal diseases globally,
putting a huge burden on the global healthcare
system. Early diagnosis of diabetes is hence, of
utmost importance and could save many lives.
However, current techniques to determine whether a
person has diabetes or has the risk of developing
diabetes are primarily reliant upon clinical
biomarkers. In this article, we propose a novel deep
learning architecture to predict if a person has
diabetes or not from a photograph of his/her retina.
Using a relatively small-sized dataset, we develop a
multi-stage convolutional neural network (CNN)-
based model DiaNet that can reach an accuracy level
of over 84% on this task, and in doing so,
successfully identifies the regions on the retina
images that contribute to its decision-making process,
as corroborated by the medical experts in the field.
This is the first study that highlights the
distinguishing capability of the retinal images for
diabetes patients in the Qatari population to the best
of our knowledge. Comparing the performance of
DiaNet against the existing clinical data-based
machine learning models, we conclude that the
retinal images contain sufficient information to
distinguish the Qatari diabetes cohort from the
control group. In addition, our study reveals that
retinal images may contain prognosis markers for
diabetes and other comorbidities like hypertension
and ischemic heart disease. The results led us to
believe that the inclusion of retinal images into the
clinical setup for the diagnosis of diabetes is
warranted in the near future.