Anomaly segmentation in retinal images with poisson-blending data augmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Diabetic retinopathy (DR) is a major diabetes complication.
• Accurate segmentation of DR lesions is crucial for early diagnosis.
• Simultaneous segmentation of multiple DR lesion types is challenging due to limited annotations and lesion diversity.

Purpose of the Study:

• To develop a novel data augmentation technique for DR lesion segmentation.
• To propose a convolutional neural network for simultaneous multi-type DR lesion segmentation.
• To improve the accuracy and efficiency of DR lesion detection and diagnosis.

Main Methods:

• Proposed a Poisson-blending data augmentation (PBDA) algorithm to generate synthetic training data.
• Developed a DSR-U-Net++ (DC-SC residual U-Net++) architecture for multi-type DR lesion segmentation.
• Conducted extensive experiments and ablation studies to validate the methods.

Main Results:

• PBDA algorithm effectively expanded training data, improving lesion segmentation.
• Ablation studies showed PBDA increased mean AUPR by >5% for all lesion types.
• DSR-U-Net++ with PBDA outperformed state-of-the-art methods on IDRiD and e-ophtha datasets.

Conclusions:

• The proposed PBDA and DSR-U-Net++ offer an efficient solution for simultaneous multi-type DR lesion segmentation.
• The developed method demonstrates significant performance improvements over existing approaches.
• This approach can be adapted for generating training data in other medical anomaly segmentation tasks.