MIDC: Medical image dataset cleaning framework based on deep learning.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computational Pathology
• Machine Learning in Healthcare

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), is vital for medical image analysis.
• High-quality datasets are essential for training accurate CNN diagnostic models.
• Mislabeled data in medical datasets significantly reduces diagnostic model performance, and non-specialist physicians struggle to identify these errors.

Purpose of the Study:

• To propose and validate a novel framework, Medical Image Dataset Cleaning (MIDC), for automatically identifying and removing mislabeled data from public medical imaging datasets.
• To enhance the accuracy and reliability of diagnostic models trained on these cleaned datasets.
• To provide a solution that does not require expert physician annotations or additional high-quality labeled datasets.

Main Methods:

• The MIDC framework utilizes multiple public datasets of the same disease, leveraging different CNNs for automated image recognition and mislabel detection.
• A novel grading rule is implemented to classify datasets into high-accuracy and low-accuracy groups.
• A CNN-based data cleaning module uses high-accuracy datasets to identify and remove mislabeled data from low-accuracy datasets.

Main Results:

• The framework was tested on four diverse medical imaging datasets: diabetic retinopathy, viral pneumonia, breast tumors, and skin cancer.
• Average diagnostic accuracy significantly increased across all tested datasets post-cleaning.
• Accuracy improvements ranged from 71.18% to 85.13% for diabetic retinopathy, 82.50% to 93.79% for viral pneumonia, 85.59% to 93.45% for breast tumors, and 84.55% to 94.21% for skin cancer.

Conclusions:

• The proposed MIDC framework effectively automates the process of cleaning mislabeled data in medical imaging datasets.
• This automated cleaning significantly enhances the diagnostic accuracy of CNN models.
• MIDC offers a valuable tool for improving the utility of public medical datasets for disease diagnosis.