Deep learning methods for mitigating catastrophic forgetting in medical imaging
Date
2025
Authors
Javadinia, Samaneh
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Continual learning allows machine learning models to learn new tasks incrementally without losing previously acquired knowledge, a capability crucial in medical imaging where data evolves over time. A persistent challenge in this field is catastrophic forgetting, where models overwrite past knowledge when learning new tasks, limiting their practical use in dynamic environments. This thesis introduces a new framework called CLFCR-MC (Continual Learning Framework with Contrastive Regularization), specifically designed to tackle catastrophic forgetting in medical imaging applications. By combining momentum contrastive learning and a custom loss function that integrates classification, cosine similarity, and distillation losses, CLFCR-MC enhances the model’s ability to retain previous knowledge while adapting to new tasks. Experiments using medical imaging datasets, such as BloodMNIST and PathMNIST, demonstrate that this framework significantly reduces forgetting and improves accuracy compared to existing methods. These findings highlight the potential of CLFCR-MC to address real-world challenges in continual learning and improve diagnostic capabilities in healthcare.
Description
Keywords
Deep learning, Continual learning, Medical imaging, Computer vision