Weakly-Supervised Learning With Complementary Heatmap for Retinal Disease Detection.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate detection of retinal diseases is vital for diagnosis.
• Convolutional Neural Networks (CNNs) with attention modules generate heatmaps for visual explanations.
• Standard heatmaps may miss crucial information in complementary regions, leading to misclassification.

Purpose of the Study:

• To develop a novel method for multi-retinal disease detection from fundus images using complementary heatmaps.
• To address challenges in fundus image analysis, including overlapping lesions and specific pathological structures.
• To improve both detection and classification accuracy for retinal diseases.

Main Methods:

• Proposed a CAM-based method for 2D color fundus images.
• Introduced two new loss functions: attention-explore loss and attention-refine loss.
• Trained the model using both accurate and inaccurate heatmaps selected based on ground truth prediction scores.

Main Results:

• The developed method demonstrated improved detection accuracy for multiple retinal diseases.
• Classification performance was enhanced concurrently with increased detection accuracy.
• Generated heatmaps more accurately highlighted lesion regions compared to existing methods.

Conclusions:

• The proposed approach effectively enhances multi-retinal disease detection and classification using complementary heatmaps.
• The novel loss functions contribute to generating more precise visual explanations for CNN models in medical imaging.
• This method offers a promising advancement for automated diagnosis of retinal conditions from fundus images.