Combating Medical Label Noise through more precise partition-correction and progressive hard-enhanced learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Deep neural networks for medical diagnosis require high-quality labeled data.
• Manual annotation of medical images can introduce label noise due to complexity and expertise required.
• Label noise can negatively impact the training and performance of classification models.

Purpose of the Study:

• To develop a noise-tolerant framework for medical image classification.
• To effectively address and mitigate the challenges posed by label noise in medical datasets.
• To improve the accuracy and robustness of computer-aided diagnosis systems.

Main Methods:

• A two-phase framework: fore-training correction and progressive hard-sample enhanced learning.
• Dual-branch sample partition detects clean, hard, and noisy instances.
• Hard-sample label refinement and joint correction enhance data quality.
• Progressive reinforcement learning improves feature representation learning.

Main Results:

• Achieved 82.39% accuracy on a pneumoconiosis dataset.
• Demonstrated robust performance on a five-class skin disease dataset with varying noise levels (up to 40%).
• Showcased high accuracy in binary polyp classification even with significant label noise (up to 40%).

Conclusions:

• The proposed framework effectively handles label noise in medical image classification.
• Demonstrated robustness and effectiveness across diverse datasets and noise levels.
• Validates the potential for improved computer-aided diagnosis systems with noisy labels.