Abstract
Purpose :
To compare the performance of three Convolutional Neural Network (CNN) models, each using a different input crop size, in detecting glaucomatous optic neuropathy (GON).
Methods :
The data used for this project were collected during an international multicentric clinical study. The dataset consisted of 1930 color retinal images including the Optic Nerve Head (ONH) automatically captured by the Compass fundus perimeter (Centervue, Italy). Each image was labelled as healthy (NRM) or glaucoma (GLC) based on a glaucoma experts’ evaluation (clinical judgment of the nerve including ophthalmoscopy, fundus image and OCT) for a total of 1010 NRM and 920 GLC. The original images were cropped and then resized to three different sizes: 200x200 (Square Crop - SC), 700x300 (Halfmoon Crop – HC), 700x500 (Fovea-Extended Crop – FEC) pixels, as reported in Figure1. The CNN models consisted of sequential convolutional and max pooling layers for a total of 1,278,113, 1,738,321 and 1,967,441 trainable parameters respectively for SC, HC and FEC models. All models had one final output node consisting of a sigmoid activation with threshold set to 0.5. The models were trained using the 90% of the dataset (80% train, 10% validation) and tested using the remaining 10% of the data randomly sampled across all study sites. The same data split was used for all models to ensure performance comparability. The metrics used to assess the performance are sensitivity, specificity, accuracy and the Area Under the ROC curve (AUC). Additionally, the 5-fold and 10-fold Cross Validation (CV) were employed to obtain more robust performance evaluations.
Results :
The results obtained on the same test set (101 NRM, 92 GLC) are reported in Table1. Additionally, AUC values are 0.956, 0.964 and 0.963 respectively for SC, HC and FEC models.
Conclusions :
All models show similar performance, as confirmed by the CV results and by the AUC values. The HC model is the only model with both sensitivity and specificity above 90% and thus it seems to be the best choice between the three models. HC model could have applications for automated case finding in glaucoma care.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.