Incremental Learning in Medical Imaging: A Comprehensive Survey of Deep Neural Network Advances and Challenges

Deep neural networks (DNNs) have achieved remarkable success in medical image analysis; however, their deployment in real-world clinical environments remains constrained by the static nature of conventional training pipelines. In practice, medical data evolve continuously as new imaging modalities, institutions, and disease categories emerge, necessitating adaptive models capable of learning incrementally without catastrophic forgetting. While lifelong learning (LL) and continual learning (CL) frameworks broadly target continuous model adaptation, incremental learning (IL) emphasizes structured, phase-wise knowledge acquisition involving new classes, domains, or modalities. Despite increasing attention to CL in computer vision, IL-specific methodologies, evaluation protocols, and benchmarks tailored to medical imaging remain largely underexplored. This paper presents the first systematic survey dedicated to incremental learning in medical imaging. A unified taxonomy is proposed across three complementary axes: (i)Learning Mechanisms— including regularization-, replay-, dynamic architecture-, parameter modulation-, and hybrid/meta-based methods; (ii)Incremental Scenarios— covering class-, domain-, task-, modality-, and data/online-incremental settings; and (iii)Application Dimensions—encompassing task types, imaging modalities, data-access constraints, and challenge-driven factors such as imbalance, forgetting, and privacy. For each axis, representative algorithms, evaluation protocols, and benchmark datasets are synthesized to illustrate how IL strategies extend continual learning principles toward clinically deployable AI systems. Finally, the survey highlights emerging research directions—such as privacy-preserving and federated IL, data-free adaptation, uncertainty calibration, and scalable integration with foundation models—laying the groundwork for robust and sustainable clinical AI development.