AlterNet-K: a small and compact model for the detection of glaucoma

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness globally.
• Early-stage glaucoma often lacks noticeable vision changes, underscoring the need for timely diagnosis.
• Manual analysis of fundus images for glaucoma detection is time-consuming.

Purpose of the Study:

• To develop a parameter-efficient deep learning model for accurate glaucoma detection using fundus images.
• To improve the generalizability of diagnostic models by combining convolutional neural networks (CNNs) and self-attention mechanisms.

Main Methods:

• A novel AlterNet-K model was proposed, integrating ResNets with multi-head self-attention (MSA).
• The model was trained and validated on the large-scale Rotterdam EyePACS AIROGS dataset (113,893 fundus images).
• Performance was compared against various transformer and CNN models.

Main Results:

• The AlterNet-K model achieved high performance metrics: 0.916 accuracy, 0.968 AUROC, and 0.915 F1 score.
• It outperformed established models like ViT, DeiT-S, Swin transformer, ResNet, EfficientNet, MobileNet, and VGG.
• Smaller CNNs integrated with MSA demonstrated superior performance over larger counterparts.

Conclusions:

• Parameter-efficient models combining CNNs and MSA can achieve high accuracy in glaucoma classification.
• The AlterNet-K model offers a promising, efficient approach for automated glaucoma diagnosis.
• This methodology can be adapted for other medical imaging classification tasks.