4S-DT: Self-Supervised Super Sample Decomposition for Transfer Learning With Application to COVID-19 Detection.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Transfer learning from large datasets improves medical image classification.
• Challenges exist in medical imaging due to data irregularity and imbalanced classes.
• Existing methods struggle with robust knowledge transfer in complex medical datasets.

Purpose of the Study:

• Propose a novel deep convolutional neural network, the 4S-DT model.
• Enhance transfer learning for chest X-ray classification using self-supervised learning.
• Improve robustness in medical image classification tasks with irregular or imbalanced data.

Main Methods:

• Developed a self-supervised super sample decomposition for transfer learning (4S-DT) model.
• Implemented a coarse-to-fine transfer learning strategy from large-scale image recognition to chest X-ray classification.
• Utilized a class-decomposition (CD) layer for downstream learning to simplify data structure and handle irregularities.

Main Results:

• The 4S-DT model achieved 99.8% accuracy on a larger dataset and 97.54% on a smaller, augmented dataset.
• Successfully applied to COVID-19 detection using 50,000 unlabeled chest X-ray images.
• Demonstrated high robustness in knowledge transformation and handling data irregularities.

Conclusions:

• The 4S-DT model offers a robust approach to transfer learning in medical image classification.
• Self-supervised sample decomposition effectively addresses data irregularities and class imbalance.
• The model shows significant promise for applications like COVID-19 detection from chest X-rays.