A Novel Framework for Sequential Multi-Source Transfer Learning: from Theory to Cancer Detection

Transfer Learning has advanced AI applications across domains like autonomous systems, natural language processing, and medical imaging. By leveraging pre-trained models, transfer learning enhances performance on small datasets. However, traditional methods – such as ensemble and multi-source transfer learning – suffer from high computational costs, memory constraints, and the need for simultaneous model access, limiting their use in resourceconstrained healthcare settings. To overcome these challenges, we propose a sequential transfer learning framework that enables incremental learning from multiple source models while reducing memory and computational demands. Unlike existing approaches, our method allows efficient fine-tuning without requiring all models to be available simultaneously. Empirical evaluations on benchmark datasets (BRACS, BACH, IDC, and Places365) using ResNet-50 and transformer-based architectures demonstrate that our approach can match or surpass the performance of traditional multi-source and ensemble methods, while improving efficiency and integrating diverse knowledge sources. By establishing key theoretical insights into multi-source and sequential transfer learning, this work advances transfer learning methodologies and their potential diagnostic accuracy in clinical applications.