Class-incremental learning using push-pull autoencoder for chest X-ray diagnosis

Class-incremental learning helps models adapt to new classes without past data. This supports long-term learning. In medical imaging, it integrates new diseases or imaging methods over time. CIL models for natural images struggle in chest x-rays diagnosis. Conversely, deep learning models that excel in chest x-ray diagnosis tend to suffer from catastrophic forgetting in a class-incremental setting. To address this, we propose a novel class-incremental learning framework specifically designed for chest x-ray analysis. Our approach utilizes both abnormality-specific and abnormality-agnostic information from the input data. We designed a model, the Push-Pull Autoencoder (PPAE), that uses a dual latent space representation to refine feature representations by disentangling abnormality-specific from abnormality-agnostic information. This enhances the model's understanding of disease features. PPAE is trained in such a way that it brings the training samples closer together based on shared, abnormality-agnostic features, while simultaneously distinguishing them using abnormality-specific features. To retain critical knowledge without exhaustive retraining, we employ a coreset generation algorithm that selects representative exemplars from previous classes. This approach maintains diagnostic accuracy on previously learned diseases while adapting seamlessly to new classes. PPAE addresses the problem of catastrophic forgetting prominent in class-incremental learning. Experimental results show up to 3% improvement in terms of F1 score and up to 4% improvement in terms of AUROC across various chest x-ray datasets. It validates the robustness of our framework in incremental learning tasks, highlighting its potential to advance continuous chest x-ray diagnosis.