Searching Discriminative Regions for Convolutional Neural Networks in Fundus Image Classification With Genetic...

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computer vision

Background:

• Deep convolutional neural networks (CNNs) excel at fundus image classification but lack explainability due to their black-box nature.
• Limited explainability hinders the clinical adoption of CNNs for medical image interpretation.

Purpose of the Study:

• To develop a novel method for identifying discriminative regions in fundus images to improve CNN explainability.
• To enhance user confidence in CNN predictions by revealing image areas crucial for classification.

Main Methods:

• Utilized an evolutionary process to select superpixels, automatically identifying discriminative regions within fundus images.
• Proposed a method to search for image regions that are most influential for a CNN's specific category classification.

Main Results:

• Demonstrated significant effectiveness in identifying discriminative regions, with an average increase of 77.8% in classification confidence.
• Showcased that identified discriminative regions align with human-interpretable evidence and are distributed across the image.
• Confirmed that a small subset of superpixels is sufficient for a CNN to confidently classify specific categories.

Conclusions:

• The proposed method effectively enhances the explainability of CNNs in fundus image classification.
• Identified discriminative regions can increase confidence in AI-driven diagnostic predictions.
• This approach aids clinicians in understanding the basis of CNN decisions in medical imaging.