Abstract
Purpose :
Image curation is one of the fundamental aspects of AI algorithm training, and application of deep learning (DL)and application of DL methods for retinal diseases depends on well-organized, representative images. In this project, we have developed a DL workflow to more efficiently curate retinal photographs, relieving the need for manual evaulation by graders.
Methods :
Stereoscopic four-field wide-capture images of the retina are commonly used for evaluating diabetic retinopathy in clinical trials. This entails acquiring red reflex images followed by stereo pair of the disc, macula, a superior, nasal, and inferior field (figure) using 50 – 60-degree camera. Graders review these images in a systematic fashion, one eye at a time, to arrive at a diabetic retinopathy severity level. Submission from clinical sites is not always organized in a manner conducive for grading or for DL training. We designed and trained a neural network to label individual image files in 3 aspects: (1) differentiating red reflex image from a retinal photograph (2) differentiating right eye images from left eye, and (3) identifying the field of the retina. Based on these three parameters, labels were generated by the model and affixed to the image file name. 1631 annotated images from 88 participants were used to train the model, and 404 images from 22 participants were used for validation during training. An independent dataset of 180 images (12 participants) with variable image quality was used for testing the performance of the trained DL model. A human grader independently reviewed the quality of field definition and assigned the field number. This was compared with the DL-generated label. All images were clinical trial submissions for various DR trials across multiple international sites.
Results :
Field definition was considered adequate quality in 141 images (78.3%). The labeling model was accurate in 152 / 180 (84.4 %) images in the testing dataset. Of the 28 inaccurately labeled images 18 (64.3%) had poor field definition.
Conclusions :
The deep learning model is an accurate automated method that can assist with workflows for organization of images for grader evaluation. This also helps develop well curated training images for future DL development towards disease identification.
This is a 2021 ARVO Annual Meeting abstract.