UniPrompt-CL: Sustainable Continual Learning in Medical AI with Unified Prompt Pools

Abstract

Modern AI models are typically trained on static datasets, limiting their ability to continuously adapt to rapidly evolving real-world environments. While continual learning (CL) addresses this limitation, most CL methods are designed for natural images and often underperform or fail to transfer to medical data due to domain bias, institutional constraints, and subtle inter-stage boundaries. We propose UniPrompt-CL, a medical-oriented prompt-based continual learning method that improves prompt pool design via a minimally expanding unified prompt pool and a new regularization term, achieving a better stability-plasticity trade-off with lower computational cost. Across two domain-incremental learning settings, UniPrompt-CL effectively reduces inference cost while improving AvgACC by 1-3 percentage points. In addition to strong performance, extensive experiments clearly validate the motivation and effectiveness of the proposed improvements.

Figures (4)

• Figure 1: To support our claim, we visualized representative examples from the natural images used in prior PCL methods and the medical datasets employed in this study.
• Figure 2: (a) visualizes the prompts learned by the independent prompt pool of OS-Prompt kim2024one and those produced by our proposed integrated pool. The dots corresponding to layers 1-5 represent the layer-wise prompts of OS-Prompt, whereas the brown dots (Ours) denote the prompts generated by our integrated pool (see the top-right legend for details). (b) further shows that, as training progresses, newly added prompts avoid redundant or overlapping representations, indicating that each prompt captures distinct and complementary features.
• Figure 3: This figure presents the overall architecture proposed in this study, which integrates an enhanced prompt pool. At each layer, the [CLS] token serves as a query, and the resulting layer-wise queries are centrally managed through the prompt pool integration module. This integration generates the prompts for the subsequent layer, which are then combined with $x_l$ and propagated forward. Additionally, at each training stage, only a small number of new prompts are introduced through a minimal prompt expansion mechanism, while all previously learned prompts remain frozen.
• Figure 4: Representative examples from the three skin cancer datasets used in our pilot study: ISIC rotemberg2021patient, HAM codella2019skin, and DERM7 kawahara2018seven. The samples illustrate the diversity in lesion appearance and acquisition conditions across datasets.