Pathological Myopia Image Recognition Strategy Based on Data Augmentation and Model Fusion.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Automatic diagnosis of retinal diseases using fundus images aids clinical decisions.
• Convolutional neural networks (CNNs) show promise but are prone to overfitting due to high expression ability.
• Traditional data augmentation (DA) methods are insufficient for modern CNN architectures with numerous parameters.

Purpose of the Study:

• To introduce a novel multimodal fusion data augmentation (DAMF) strategy to enhance CNN performance in retinal disease diagnosis.
• To address the limitations of traditional DA techniques in enriching datasets for complex CNN models.
• To improve the generalization ability of diagnostic models through optimized data and classifier fusion.

Main Methods:

• Proposed a multimodal fusion data augmentation (DAMF) strategy integrating standard DA, data disrupting, data mixing, and autoadjustment methods.
• Utilized DAMF to enhance image data in the training dataset, creating new training images.
• Implemented classifier result fusion via voting based on DAMF to improve model generalization.

Main Results:

• The DAMF strategy successfully identified optimal data augmentation modes for specific image datasets.
• Evaluation on the iChallenge-PM dataset demonstrated DAMF's effectiveness.
• The optimal DAMF configuration achieved a 2.85% accuracy increase compared to the original training dataset.

Conclusions:

• The proposed DAMF strategy effectively enhances CNN performance for retinal disease diagnosis.
• DAMF offers a flexible approach to optimize data augmentation for diverse datasets.
• This method improves model accuracy and generalization, supporting clinical decision-making in ophthalmology.