Task Augmentation-Based Meta-Learning Segmentation Method for Retinopathy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) for medical image segmentation demands extensive labeled data, which is costly and time-consuming to acquire.
• Existing meta-learning approaches lack sufficient multi-task datasets for rich-class medical image segmentation.
• Variations in data from different sources can degrade model performance.

Purpose of the Study:

• To develop a task augmentation-based meta-learning method (TAMS) for efficient retinal image segmentation.
• To address the labor-intensive annotation demand in medical image analysis.
• To create diverse, high-quality medical image datasets without extensive manual labeling.

Main Methods:

• Proposed a Retinal Lesion Simulation Algorithm (LSA) for automatic generation of multi-class retinal disease datasets with pixel-level labels.
• Designed a novel simulation function library to control the generation process and ensure interpretability.
• Introduced a generative simulation network (GSNet) with enhanced adversarial training for high-quality representations of complex retinal diseases.

Main Results:

• TAMS demonstrated superior segmentation performance compared to state-of-the-art models.
• The LSA successfully augmented meta-learning tasks without requiring new data collection.
• GSNet maintained high-quality representations for complex retinal pathologies.

Conclusions:

• The proposed TAMS method effectively addresses the challenge of limited labeled data in medical image segmentation.
• Automated data generation via LSA and GSNet offers a scalable solution for creating diverse medical imaging datasets.
• TAMS provides a significant advancement in automated retinal image segmentation, outperforming existing methods.