Multiple Lesions Insertion: boosting diabetic retinopathy screening through Poisson editing.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Collecting large-scale medical image datasets with ground truth is challenging due to high professional requirements.
• Deep neural networks in computer vision tasks often require extensive data, which is difficult to obtain for medical imaging.

Purpose of the Study:

• To propose a novel data augmentation method, Multiple Lesions Insertion (MLI), for synthesizing realistic diabetic retinopathy (DR) fundus images.
• To enhance the performance of computer-aided diagnosis (CAD) systems for DR screening by addressing data scarcity.

Main Methods:

• The MLI method synthesizes new DR fundus images by inserting real lesion templates (exudates, hemorrhages, microaneurysms) into healthy fundus images using Poisson editing.
• Synthetic images are generated adhering to clinical rules regarding lesion distribution across different DR grades.
• The feasibility of MLI was demonstrated within a DR computer-aided diagnosis (CAD) system for treatment transfer judgment.

Main Results:

• Generated DR fundus images exhibit realistic textures and rich details, free from artifacts and discontinuities.
• The MLI method demonstrated superior performance in DR screening tasks compared to traditional augmentation techniques like oversampling, undersampling, cropping, and rotation.
• The synthesized images effectively supplement insufficient medical image datasets.

Conclusions:

• The MLI method is a feasible and effective approach for generating realistic synthetic DR fundus images.
• MLI significantly improves the performance of DR screening tasks, outperforming conventional data augmentation strategies.
• This technique offers a valuable solution for augmenting limited medical image datasets, particularly for training AI models in healthcare.