Medical image perception is crucial for clinical diagnostics and treatment planning, yet faces challenges in modeling long-range dependencies, computational efficiency, and adapting to evolving clinical needs without catastrophic forgetting. To address these issues, we propose a Semantic Enhancement Network (SENet), which integrates Mamba-based architecture with tailored continual learning strategies. It introduces a Semantic Enhancement framework through several key innovations: an Orthogonal Gradient Correction Module (OGCM) and task-specific prompts to mitigate catastrophic forgetting at both image and sequence levels, enabling effective continual learning, an Attention-guided State Space Layer (ASSL) that captures comprehensive semantic features, and a Local Semantic Attention (LSA) module to refine local detail extraction. These components collectively enhance feature fusion by adaptively integrating multi-scale and contextual information. Extensive experiments on LIDC and LUNA16 lung nodule perception datasets demonstrate that SENet achieves state-of-the-art performance, with notable improvements over strong baselines, e.g., gains of 0.22%–0.59% in mDice and 1.43%–2.26% in mMPA, while maintaining the lowest computational cost (1.12G FLOPs) and memory usage (2.82MB), validating its precision, efficiency, and generalization capability for dynamic clinical environments.
Jin et al. (Fri,) studied this question.