Abstract
Purpose :
To develop a deep learning algorithm to detect glaucoma based on fundoscopy and try to validate it with visualization methods
Methods :
We retrospectively collected color fundoscopic pictures from National Taiwan University Hospital. The inclusion criteria of the glaucoma group were defined as first-diagnosed POAG and PACG patients. Secondary glaucoma was excluded. The control group consisted of ocular hypertensive patients. A total of 67 patients and 56 patients were in glaucoma and control group, respectively. Besides, we also included 67 healthy color fundoscopic photos from the public dataset for training. After data augmentation, we adopted the modified Oxford VGG16 model for learning under 10-fold cross-validation. The automatic optic rim delineation was incorporated into the algorithm for increasing the accuracy of the model. The sensitivity, specificity, and area under receiver operating characteristic curve (AUROC) were calculated. For better explainability, we displayed the pixel-wise discriminative features and class-discriminative heat map of diopter images for visualization.
Results :
A total of 207 glaucoma pictures and 200 normal pictures were collected from National Taiwan University Hospital and 67 normal pictures from the public database. After training for 1000 epoch, the convergence of accuracy and loss for training and validation datasets which showed no overfitting. The sensitivity, the specificity, and the AUROC were 0.90, 0.895 and 0.944, respectively. The pixel-wise discriminative features visualized by the guided backpropagation and the heat map of the prediction layer by the guided grad-CAM all showed the deep learning model focused on the optic nerve, which was corresponding to the diagnostic clues of the ophthalmologists. T
Conclusions :
In this study, the CNN model had fair accuracy for glaucoma screening based on the color funduscopic images. The visualization mentioned in this study revealed the model focus on the proper region for diagnosis and made clinical explainability of deep learning more acceptable.
This is a 2020 ARVO Annual Meeting abstract.