Abstract
Purpose :
Interpretability is the key factor in automatic classification in medical diagnosis design. The deep Learning (DL) methodology, and its subfield Convolutional neural networks (CNN), takes in an input image, identifies some aspects in the picture, wich are important to define it, and using parametric models classificate and differenciate it from others . The novelty of this desing from the previous is its ability to learn theese filters with the experience.
Diabetic Retinopathy (DR) is a leading disabling chronic disease and one of the main causes of blindness and visual impairment in developed countries. It is reported that a 90% of the cases DR can be prevented through early detection and treatment. It is developed an automated learning algorithm model for the DR to assist physiscian in DR screening.
Methods :
We validated our algorithm of automated learning diabetic retinopathy (ALDR-algorithm), made by CNN for the imaging interpretation. It was tried on 56.088 database images, and It was tested in 1.748 retinographies from Messidor-2 database. ALDR-algorithm was also tried on 38.319 images from our own population and it was tested in 5000 of these retinographies. The screening was made by a single retinography from both eyes acording EURODIAB recomendations. To test the algorithm concordance with the experts, the images were read by our ALDR-ALGORITHM and by four masked retinologists. Retinopathy was graded according to MESSIDOR in four levels. The agreement, sensitivity and specificity between both methods was determined.
Results :
The agreement between both models was a QWK index of 0,918.
The ALDR-algorithm statistical results of the images with any DR were: sensitivity =0.985, specificity =0.853, positive predictive value =0.980, negative predictive value =0.980, false positive rate or error type I = 0,147, false negative rate or error type II=0,015.
The results for referable DR images (moderate, severe or proliferative DR) were: sensitivity =0.991, specificity =0,850, positive predictive value =0,987, negative predictive value =0,888, false positive rate or error type I = 0,150, false negative rate or error type II=0,009.
Conclusions :
It has been developed an ALDR-algorithm able to read retinographies of diabetic patients with a great agreement, and it would be useful in DR screening.
This is a 2020 Imaging in the Eye Conference abstract.