125 — A Deep Learning Interpretable Classifier for Diabetic Retinopathy Disease Grading

de la Torre et al (1712.08107)

Read on 23 December 2017
#retina  #retinopathy  #diabetes  #diabetic-retinopathy  #dnn  #neural-network  #explainability  #EyePACS 

In this paper, de la Torre et al propose a deep neural network model that classifies retinal scans based on a label of “Non-Proliferative Diabetic Retinopathy”, or (NPDR) — classifying as Mild, Moderate, Severe, or — a final category — Proliferative Diabetic Retinopathy (PDR).

Diabetic retinopathy is a common disease state in which high blood sugar levels damage the fragile blood vessels in the retina, causing swelling, leaking, or flow-blockage. It is a very common cause of onset blindness.

This network is unique among DR detection networks in its emphasis on explainability, a feature uncommon in neural networks in general but a highly sought-after feature in radiology and medical-imaging algorithms. The network both returns a global classification (proliferative or non-proliferative, as well as a level of severity) as well as a pixelmap indicating which pixels were most influential in deciding upon the classification.

This can be overlaid on the retinal scan in order to direct the attention of radiologists when confirming the diagnosis, or it can be used as a training layer for a future neural network, with corrections and human annotations used to recalibrate the net.

This network achieves expert-level classification across the five levels of diagnostic classes, and demonstrates a proof of concept for better explainability in future radiologist networks.