Robust co-teaching learning with consistency-based noisy label correction for medical image classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks (DNNs) excel in computer-aided diagnosis but require high-quality labeled data.
• Noisy labels in medical datasets, arising from observer variability, significantly hinder DNN performance.
• Existing methods struggle with accurate noise rate estimation in clinical settings.

Purpose of the Study:

• To propose a robust noisy label correction method using a co-teaching paradigm for medical image datasets.
• To address the performance degradation of DNNs caused by noisy labels in medical imaging.
• To develop a method that can effectively correct or remove noisy labels, improving classification accuracy.

Main Methods:

• A co-teaching learning paradigm is employed for robust noisy label correction.
• An adaptive noise rate estimation module calculates the dataset's noise rate.
• A consistency-based noisy label correction module identifies and corrects noisy labels.

Main Results:

• The proposed method demonstrated superior performance in medical image classification tasks on ISIC-2017, ISIC-2019, and a thyroid ultrasound dataset.
• The method was also evaluated on the CIFAR-10 natural image dataset, showing good generalization capabilities.
• Experimental results confirm the effectiveness of the noisy label correction approach.

Conclusions:

• The proposed method effectively handles noisy labels in medical image datasets through self-adaptive correction.
• The technique is suitable for medical image classification, improving the reliability of DNNs.
• This approach offers a valuable solution for leveraging imperfectly labeled medical imaging data.