Understanding and interpreting CNN's decision in optical coherence tomography-based AMD detection.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Automated assessment of age-related macular degeneration (AMD) using optical coherence tomography (OCT) is a growing research area.
• Existing convolutional neural network (CNN) methods for AMD detection lack interpretability.
• This study addresses the need for understanding CNN decision-making processes in AMD diagnosis.

Purpose of the Study:

• To investigate the decision-making processes of CNNs used for AMD detection.
• To identify specific retinal layers influencing CNN predictions.
• To bridge the gap in interpreting CNNs' diagnostic decisions in OCT scans.

Main Methods:

• Trained multiple CNN models (VGG16, VGG19, Xception, ResNet50, InceptionResNetV2) for AMD detection.
• Applied CNN visualization techniques (Grad-CAM, Grad-CAM++, Score CAM, Faster Score CAM) to identify regions of interest.
• Developed retinal layer segmentation to correlate CNN focus with retinal structures using 2130 SD-OCT scans.

Main Results:

• The Outer Nuclear Layer to Inner Segment Myeloid (ONL-ISM) significantly influences AMD detection decisions across all tested CNN models.
• Normalized Intersection (NI) scores indicated varying degrees of ONL-ISM influence for AMD versus normal cases.
• Specific NI scores were reported for each CNN model, demonstrating differential contributions of ONL-ISM.

Conclusions:

• The ONL-ISM is the primary retinal layer contributing to CNN-based AMD detection.
• The Nerve Fiber Layer to Inner Plexiform Layer (NFL-IPL) is identified as the second most influential layer.
• This research provides critical insights into the interpretability of CNNs for AMD diagnosis.