POAGNet

POAGNet employs a Siamese neural network to analyze paired optic disc photographs for detecting and monitoring primary open-angle glaucoma (POAG).

Key Features:

• Siamese Network Architecture: Processes pairs of optic disc photographs to assess changes over time, mimicking clinical comparison of baseline and follow-up images.
• Deep Supervision: Incorporates two side outputs for deep supervision to enhance training efficiency and accuracy.
• Convolutional Similarity Measurement: Uses convolution operations to measure similarity between paired images for incident glaucoma assessment.
• Training Datasets: Trained and evaluated on the Ocular Hypertension Treatment Study (OHTS) dataset (37,339 optic disc photographs from 1,636 participants) and the Sequential fundus Images for Glaucoma (SIG) dataset (3,684 optic disc photographs).
• Performance Metrics: Evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity, achieving AUCs of 0.9587 (OHTS test set) and 0.7518 (SIG test set).
• Generalizability and Transferability: Demonstrated transferability with OHTS-trained model tested on SIG (AUC 0.7490) and improved performance when trained on combined SIG+OHTS (AUC 0.8165).

Scientific Applications:

• Early detection and diagnosis of POAG: Automated analysis of sequential optic disc photographs to detect incident primary open-angle glaucoma.
• Longitudinal monitoring of glaucomatous progression: Assessment of changes between baseline and follow-up optic disc images to support progression evaluation.
• Cross-dataset validation and transfer learning assessment: Evaluation of model performance and transferability across OHTS and SIG datasets and combined datasets.

Methodology:

Computational methods explicitly stated include a Siamese neural network processing paired optic disc photographs, two side outputs for deep supervision, convolutional operations to measure pairwise image similarity, and training/evaluation on the OHTS and SIG datasets with performance measured by AUC, accuracy, sensitivity, and specificity.