Confident Learning-Based Label Correction for Retinal Image Segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Accurate labeling is critical in medical image analysis, especially for diabetic retinopathy detection.
• Label noise in retinal images can lead to diagnostic errors and complicate segmentation tasks.

Purpose of the Study:

• To develop and evaluate an innovative label correction framework for pixel-level inaccuracies in retinal image segmentation.
• To combine Confident Learning (CL) with human-in-the-loop re-annotation to enhance data quality.

Main Methods:

• Two CL strategies, Confident Joint Analysis (CJA) and Prune by Noise Rate (PBNR), were assessed using DeeplabV3+ and ResNet architectures.
• The methods were applied to four public datasets (HRF, STARE, DRIVE, CHASE_DB1) to quantify and correct label noise.
• Performance was evaluated based on accuracy, Intersection over Union (IoU), and Mean Boundary F1 Score (MeanBFScore) before and after noise reduction.

Main Results:

• Label noise reduction consistently improved accuracy, IoU, and weighted IoU across all datasets.
• Segmentation of small structures like the fovea showed significant enhancement after refinement.
• CL with human refinement boosted segmentation accuracy and evaluation robustness, achieving notable increases in key metrics.

Conclusions:

• The proposed methodology offers a feasible and scalable solution for addressing label noise in medical image analysis.
• This approach holds significant potential for improving the accuracy and reliability of automated diagnostic tools in clinical settings.
• The framework enhances segmentation accuracy and evaluation robustness, crucial for real-world applications.