Laterality Classification of Fundus Images Using Interpretable Deep Neural Network.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated classification of fundus images is crucial for efficient clinical workflows.
• Convolutional neural networks (CNNs) show promise in medical image analysis.
• Understanding model outputs, including feature importance and uncertainty, is vital for clinical adoption.

Purpose of the Study:

• To analyze the outputs of a CNN model for classifying fundus image laterality.
• To enhance the interpretability of the CNN model through visualization and uncertainty quantification.
• To assess the model's performance and its potential benefits for clinicians.

Main Methods:

• A CNN model was trained and tested on 25,911 fundus images.
• Activation maps were generated to identify important image regions for classification.
• Uncertainty quantification was employed to analyze misclassifications.

Main Results:

• The model achieved a mean training accuracy of 99%, comparable to clinicians.
• High activation was observed at the optic disc and surrounding retinal blood vessels, consistent with clinical attention.
• Misclassified images were associated with high prediction uncertainties and were often ungradable.

Conclusions:

• The developed CNN model effectively classifies fundus image laterality with high accuracy.
• Visualization of informative regions and uncertainty estimation improve model interpretability for clinicians.
• The system offers a potential tool to aid clinicians in fundus image analysis and reduce their workload.