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

Meta-learning-based models trained on multiple classification tasks based on multiple classes can adapt to new classification tasks with limited training samples, thereby achieving few-shot learning. However, when the number of classes in the classification task chest X-ray image analysis is also limited, meta-learning can result in overfitting. This study sought to overcome this with a class augmentation method using a generative adversarial network to generate pseudo-classes, thereby increasing the number of classes. The proposed method was evaluated using three two-way classification tasks (chronic obstructive pulmonary disease (COPD) vs. non-COPD; atelectasis vs. pneumothorax (PX); and tuberculosis (TB) vs. nontuberculous mycobacteria) and one three-way classification task (atelectasis vs. PX vs. pneumonia). Compared to meta-learning without class augmentation, the proposed scheme increased the accuracy of the two-way 50-shot tasks by 7.14%, 4.47%, and 4.43%, respectively. The proposed method also increased the accuracy of the three-way 50-shot classification task by 2.5%. This suggests potential in reducing image labeling needs and training models for rare diseases.