Understanding How Fundus Image Quality Degradation Affects CNN-based Diagnosis

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Clinical fundus images often suffer from quality degradation.
• Convolutional Neural Network (CNN) models are widely used for retinal fundus image interpretation.
• The performance of CNNs under quality degradation (QD) is not well understood.

Purpose of the Study:

• To systematically investigate the effects of QD on CNN-based diagnostic models for diabetic retinopathy (DR).
• To analyze the impact of specific quantified interferences on DR grading systems.
• To evaluate the robustness of various CNN architectures under controlled image degradation.

Main Methods:

• Controlled introduction of quantified interferences (blurring, artifacts, light disturbance) to fundus images.
• Evaluation of multiple CNN models (AlexNet, SqueezeNet, VGG, DenseNet, ResNet) on degraded images.
• Analysis of DR grading performance based on diagnosis accuracy on impaired images.

Main Results:

• Image blurring significantly reduces CNN model performance.
• Light transmission disturbance and retinal artifacts have a relatively minor impact on diagnostic accuracy.
• VGG, DenseNet, and ResNet demonstrated superior performance and robustness against degradation.

Conclusions:

• Image blurring is a critical factor affecting CNN-based DR diagnosis.
• Certain CNN architectures (VGG, DenseNet, ResNet) exhibit better resilience to common fundus image quality issues.
• Understanding QD effects is crucial for reliable AI-powered ophthalmic diagnostics.