Robust Medical Image Classification From Noisy Labeled Data With Global and Local Representation Guided Co-Training

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks (DNNs) excel in image analysis but require accurate annotations.
• Noisy labels in medical image datasets hinder DNN performance due to reliance on expert annotators.
• High-quality annotated medical data is scarce, posing a significant challenge.

Purpose of the Study:

• To develop a robust medical image classification method for handling noisy-labeled data.
• To improve the accuracy and reliability of deep learning models in medical image analysis despite data imperfections.
• To address the challenge of limited high-quality annotated medical data.

Main Methods:

• A novel collaborative training paradigm integrating global and local representation learning.
• Utilizing a self-ensemble model with a noisy label filter to identify and separate clean from noisy samples.
• Implementing a collaborative training strategy for clean samples and a self-supervised approach for noisy samples.

Main Results:

• The proposed method demonstrates superior performance compared to existing noisy label learning techniques.
• Robust classification accuracy was achieved across four public medical image datasets with diverse noise types.
• Extensive experiments validated the effectiveness of individual components within the proposed strategy.

Conclusions:

• The developed collaborative training paradigm offers a robust solution for medical image classification with noisy labels.
• The global and local representation learning scheme effectively regularizes networks to utilize noisy data.
• This work contributes to overcoming data quality limitations in medical AI by enabling learning from imperfectly annotated datasets.