Related Experiment Video
Updated: Sep 20, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
Bayesian statistics-guided label refurbishment mechanism: Mitigating label noise in medical image classification
Mengdi Gao1,2,3,4, Ximeng Feng1,2,3,4, Mufeng Geng1,2,3,4
1Department of Biomedical Engineering, College of Future Technology, Peking University, Beijing, China.
Medical Physics
|June 9, 2022
Summary
This study introduces a Bayesian statistics-guided label refurbishment mechanism (BLRM) to improve deep neural network performance in medical image classification by correcting noisy labels. BLRM effectively mitigates label noise, enhancing model robustness and accuracy in medical image analysis.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Computer Vision
Background:
- Deep neural networks (DNNs) excel in medical image classification but require extensive labeled data.
- Label noise is a common issue in medical image datasets, degrading model performance.
- Robust training strategies are crucial to overcome the challenges posed by label noise in medical image classification.
Purpose of the Study:
- To develop a novel mechanism for mitigating label noise in medical image classification tasks.
- To enhance the robustness of deep learning models against noisy labels.
- To improve the overall performance of DNNs in medical image classification.
Main Methods:
- Proposed a Bayesian statistics-guided label refurbishment mechanism (BLRM) for DNNs.
- Utilized maximum a posteriori probability and exponentially time-weighted techniques for selective label correction.
- Gradually purified training images throughout the training epochs to improve classification outcomes.
Main Results:
- BLRM selectively refurbishes noisy labels, effectively reducing the impact of erroneous data.
- Experiments on synthetic and real-world noisy datasets demonstrate BLRM's efficacy.
- BLRM integrated with DNNs shows consistent performance across various noise levels and is independent of backbone architecture.
Conclusions:
- The proposed BLRM is highly capable of mitigating label noise in medical image classification.
- BLRM offers a superior approach compared to existing anti-noise methods.
- This mechanism enhances the reliability and accuracy of DNNs in medical image analysis.
