Abstract
Purpose
Using convolutional neural networks (CNNs), we attempted to discriminate right and left fundus images of the retinal nerve fiber layer (RNFL), blue autofluorescence (BAF), and infrared reflectance (IR).
Methods
We prepared sets of 36,169 RNFL images, 4,695 BAF images, and 4,420 IR images. We evaluated each image set with three tests. Test 1 compared unmodified right and left fundus images. Test 2 compared right and flipped left images. Test 3 compared only left images that were divided randomly into two subsets.
Results
In Test 1, CNNs showed high accuracy for the RNFL, BAF, and IR sets (accuracy 100%, 99.74%, and 100%, respectively). In Test 2, the RNFL and IR sets showed high accuracy (97.93% and 95.84%, respectively), while the BAF set had relatively low accuracy (66.15%). In Test 3, the CNNs did not classify the images correctly.
Conclusion
We confirmed that CNNs could distinguish monochromatic images of the right and left fundus, even after horizontal flipping. This asymmetry could result in bias in CNN models. Therefore, asymmetry between the right and left fundus should be considered when developing a CNN model.