Enhancing Few-Shot Chest X-ray Classification through Generative Class Augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Meta-learning models excel at few-shot learning but can overfit when class numbers are limited in chest X-ray analysis.
• Limited training data and few classes pose a challenge for accurate classification in medical imaging tasks.

Purpose of the Study:

• To address meta-learning overfitting in limited-class chest X-ray analysis.
• To enhance few-shot learning capabilities by increasing the number of available classes through augmentation.

Main Methods:

• A class augmentation strategy using generative adversarial networks (GANs) to create pseudo-classes.
• Evaluation on binary classification tasks (COPD vs. non-COPD, atelectasis vs. pneumothorax, tuberculosis vs. nontuberculous mycobacteria) and a ternary classification task (atelectasis vs. pneumothorax vs. pneumonia).

Main Results:

• The proposed class augmentation method significantly improved accuracy in 50-shot, two-way classification tasks compared to standard meta-learning.
• Accuracy gains of 7.14%, 4.47%, and 4.43% were observed for the binary tasks.
• A 2.5% accuracy increase was achieved for the three-way 50-shot classification task.

Conclusions:

• Class augmentation with GANs effectively mitigates overfitting in meta-learning for limited-class medical image analysis.
• The method shows promise for reducing the burden of image labeling and improving diagnostic models for rare diseases.