CNNs for automatic glaucoma assessment using fundus images: an extensive validation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Current glaucoma assessment algorithms rely on handcrafted features, limiting performance.
• Convolutional neural networks (CNNs) excel at learning discriminative features directly from image data.

Purpose of the Study:

• To evaluate the efficacy of ImageNet-trained CNN models for automatic glaucoma detection using fundus images.
• To compare the performance of different CNN architectures against existing state-of-the-art methods.

Main Methods:

• Employed five ImageNet-trained CNN models: VGG16, VGG19, InceptionV3, ResNet50, and Xception.
• Validated models using extensive cross-validation and cross-testing strategies on public and a new clinical database.

Main Results:

• The Xception architecture achieved an average AUC of 0.9605, with high specificity (0.8580) and sensitivity (0.9346).
• Significantly improved performance compared to existing state-of-the-art methods.
• Introduced ACRIMA, the largest publicly available clinical database for glaucoma diagnosis.

Conclusions:

• ImageNet-trained CNN models provide a reliable and effective approach for automated glaucoma screening.
• Publicly available models, weights, and software facilitate further research and development in glaucoma diagnosis.