A convolutional neural network (CNN) model performed as well as clinicians in classifying the area of concern in retinal fundus images and provided evidence for why those choices were made—a common problem for artificial intelligence (AI) technology.
Retinal fundus images are used to determine vision-threatening conditions. But up to this point, they have been performed manually with pre-determined rules at the location of optic disc and surrounding retinal blood vessels, authors wrote in research published in the Journal of Digital Imaging—a straightforward but “mundane” task clinicians do not like, they added.
“We expect that our model not only improves the efficiency of fundus laterality classification in clinics by delivering prompt and automatic predictions with high accuracy but also provides promising ways to [interact] with an automated system for clinicians by presenting determinant regions for the decision and estimating uncertainty in the decision,” wrote first author Yeonwoo Jang, with the department of statistics at the University of Oxford and colleagues.
The group’s model was trained and tested using 25,911 images. Of those, 43.4 percent were macula-centered images and 28.3 percent were superior and nasal retinal fundus images.
The CNN demonstrated a mean 99 percent accuracy, which authors noted was comparable to a clinician.
Prior to testing the new technique, Jang and colleagues also generated activation maps that allowed the team to visualize important regions used by the neural network when making its laterality classification decision.
The activation maps indicated optic disc and the surrounding blood vessels were used to reach a decision—a finding the team described as interesting because clinicians use the same two determinants in their classification process.
The model was also able to perform uncertainty analysis for the decisions that discovered misclassified images tended to have a higher uncertainly than correctly classified images. Those that were wrongly classified consisted overwhelmingly of images that were unusable.
“We believe that visualization of informative regions and the estimation of uncertainty, along with presentation of the prediction result, would enhance the interpretability of neural network models in a way that clinicians can be benefited from using the automatic classification system,” authors wrote.