Evaluating the diagnostic generalizability of deep learning models trained with images of the ganglion cell layer of early glaucoma

Abstract

PurposeTo determine the diagnostic generalizability of two deep learning models when trained only with images of the ganglion cell layer (GCL) of mild glaucoma.MethodsWe have collected a sample from patients with primary and secondary open‐angle glaucoma and normal patients. The sample was divided into mild glaucoma (MD≤6 dB), and moderate‐advanced (MD > 6 dB). The GCL images were recorded with a spectral‐domain Optical Coherence Tomography. Two pre‐trained models were used, the ResNet101 and the Shufflenet. The sensitivity, specificity, diagnostic precision in training and test, and the ROC area were calculated for the two models with three different training conditions according to how the images were partitioned into training and test. In the first partition, mild glaucomas were used for training and moderate‐advanced for test. In the second, moderate‐advanced glaucomas were used for training and mild for test. In the third, the whole sample was used without classifying by severity. Gradient‐weighted Class Activation Mapping (GradCAM) was used to obtain saliency maps which highlight the more important components in the images for the model prediction. The correlation coefficient between the maps of the glaucoma and normal images of the two models was calculated.Results561 eyes were collected from 337 patients, 356 are glaucomatous and 200 are normal. The precision of the models in the test set in partition 1, was 90.9% (ResNet101) and 94.2% (Shufflenet). In partition 2, was 74.4% (ResNet101) and 73.5% (Shufflenet), and in partition 3 an accuracy of 94.6% was found with both models. The correlation coefficient between the GradCAM saliency maps of the models was 0.46 for glaucoma images and 0.83 for normal images.ConclusionsThe two deep learning models are able to generalize and have high diagnostic precision if they are trained only with images of the GCL of mild glaucoma. Both models show high correlation in the GradCAM saliency maps with normal images.