Asymmetry between right and left optical coherence tomography images identified using convolutional neural networks

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Previous studies identified ocular asymmetries in fundus photographs.
• The macula region proved highly discriminative for differentiating left and right fundus images.

Purpose of the Study:

• To investigate the capability of convolutional neural networks (CNNs) in discriminating between left and right eye optical coherence tomography (OCT) images.
• To support prior findings on ocular asymmetries using OCT data.

Main Methods:

• Utilized a dataset of 129,546 OCT images.
• Employed various CNN architectures (DenseNet121, ResNet50, VGG19) for image classification.
• Assessed classification accuracy with original and flipped images.

Main Results:

• CNNs achieved high accuracy (99.50%) in distinguishing right and left horizontal OCT images.
• Discrimination remained effective even after flipping left images (90.33% for DenseNet121).
• CNNs showed lower accuracy for vertical images (86.57%) and struggled with right horizontal images (50.82%).

Conclusions:

• CNNs demonstrate significant ability to differentiate ocular laterality from OCT images.
• Image flipping can introduce bias, impacting machine learning model performance.
• Careful consideration of image orientation and augmentation is crucial for reliable AI in ophthalmology.