Co-Correcting

Co-Correcting mitigates label noise in deep learning-based medical image classification by performing dual-network mutual learning, estimating label probabilities, and applying curriculum-based label correction to improve label accuracy and classification performance.

Key Features:

• Dual-Network Mutual Learning: Employs a dual-network architecture where two networks train simultaneously and provide reciprocal feedback for cross-validation and label correction.
• Label Probability Estimation: Estimates the probability of each label being correct to prioritize learning from more reliable examples.
• Curriculum Label Correcting: Applies a curriculum learning strategy that progressively refines labels by addressing easier-to-correct instances first and then more complex cases.
• Empirical Validation: Validated on two representative medical image datasets and the MNIST dataset, demonstrating superior accuracy and generalization compared to six state-of-the-art learning-with-noisy-labels methods across varying noise ratios.

Scientific Applications:

• Medical image analysis: Improves deep learning model performance for tasks such as disease detection, diagnosis, and treatment planning when annotated data are scarce or noisy due to expert-dependent labeling.

Methodology:

Two networks are trained simultaneously with mutual learning; per-label probability estimation is used to assess label reliability; a curriculum-based, staged label correction scheme refines labels from simpler to more complex instances.