A Peer-to-Peer Federated Incremental Learning Framework for Multi-Institutional Low-Dose CT Denoising

Deep learning techniques have been widely used in computed tomography (CT), but they require large datasets to train networks. Moreover, data sharing among multiple institutions is limited due to data privacy constraints, which hinders the development of high-performance DL-based CT imaging models that rely on multi-institutional collaborations. Federated learning (FL) is an alternative strategy for training models without centralized data across multiple institutions. In this work, we propose a novel peer-to-peer federated incremental learning strategy to improve low-dose CT imaging performance across multiple institutions. The newly proposed method is called peer-to-peer incremental FL with intermediate controllers, i.e., icP2P-FL. Specifically, unlike conventional FL models, the proposed icP2P-FL does not require a central server to coordinate training information for a global model. In the icP2P-FL method, peer-to-peer federated incremental learning is introduced, where each institution jointly maintains an institution-shared sub-model with prompt learning, which shares network parameters exclusively with each decoder. The framework incorporates decoded features from the CT images of all institutions via peer-to-peer cyclic information exchange. Furthermore, an intermediate controller with self-aware regulation is designed to enhance the flexibility of overall training. Extensive experiments were conducted on the AAPM Low-Dose CT Grand Challenge dataset and local datasets. The results demonstrate that the proposed icP2P-FL method outperforms other comparative approaches both qualitatively and quantitatively, achieving accuracy comparable to that of a model trained on centralized data from all institutions.