Validation of a Deep Learning Model to Screen for Glaucoma Using Images from Different Fundus Cameras and Data...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Glaucoma diagnosis relies heavily on interpreting fundus photography.
• Variations in fundus cameras and imaging protocols across institutions can affect diagnostic consistency.
• Deep learning algorithms offer potential for automated and standardized glaucoma detection.

Purpose of the Study:

• To validate a deep residual learning algorithm for glaucoma diagnosis.
• To assess the algorithm's performance using fundus images from diverse cameras and institutes.
• To evaluate the impact of image augmentation on diagnostic accuracy.

Main Methods:

• A Residual Network (ResNet) deep learning algorithm was trained on a dataset of 1364 glaucomatous and 1768 non-glaucomatous fundus images.
• Two independent testing datasets, comprising images from different camera manufacturers and institutions, were used for validation.
• Image augmentation techniques were applied to artificially increase the training dataset size; diagnostic accuracy was measured by the area under the receiver operating characteristic curve (AROC).

Main Results:

• The algorithm achieved high diagnostic performance across both testing datasets.
• When image augmentation was utilized, the AROC reached 94.8% and 99.7% for the respective datasets.
• Without image augmentation, AROC values were significantly lower (87.7% and 94.5%), highlighting the benefit of augmentation.

Conclusions:

• The deep residual learning algorithm demonstrates high diagnostic performance for glaucoma detection.
• The algorithm's effectiveness is maintained across different fundus cameras and multiple institutions.
• Image augmentation is a crucial technique for enhancing the algorithm's diagnostic accuracy and generalizability.