Continual Medical Image Segmentation Based on Generative Replay

Cardiac magnetic resonance (CMR) imaging plays a crucial role in clinical diagnosis for heart function evaluation and cardiovascular disease detection. Although deep learning-based medical image segmentation models have made significant progress in specific scenarios, they rely on static data distributions and struggle to handle dynamically changing clinical data, limiting the generalization ability of the model and making it susceptible to catastrophic forgetting. To address this issue, this paper proposes a combination of generative models and a continual learning framework, generating synthetic samples containing historical task features for data augmentation. Additionally, low-rank regularization and style-sharing parameter compression methods are employed to optimize storage requirements and improve the long-term stability and generalization ability of model. Experiments demonstrate the effectiveness of this approach on the cardiac MRI and optic disc datasets.