Abstract: To address the challenges of scarce labeled data, cumulative distribution discrepancies, and catastrophic forgetting caused by feature drift at both domain and class levels in semi-supervised domain-incremental learning for medical image segmentation, it is essential to explore effective methods capable of continuously adapting to new domains under limited annotation conditions while preserving existing segmentation capabilities. To tackle these issues, we propose a dual-granularity prototype contrastive learning approach. First, optimal transport-driven semi-supervised feature alignment is employed to reconcile distribution differences. Second, prototype contrastive learning dynamically adapts to feature distribution shifts during the domain-incremental learning process. Finally, a historical prototype inference correction mechanism is introduced to leverage historical prototype information for refining model predictions, effectively resolving the conflicts among sparse annotation resources, dynamic data distributions, and segmentation objectives in incremental learning. Experiments on cardiac and prostate segmentation datasets demonstrate that under sparse annotation conditions, the proposed method exhibits continuous adaptation capability to new domains, stable sequential segmentation performance, and certain anti-forgetting properties. The results indicate that compared to baseline methods, the proposed approach achieves overall Dice Similarity Coefficients of 87.13% and 86.05% across multiple incremental stages on the cardiac and prostate datasets, respectively, with improvements of 5.98 and 18.64 percentage points over the suboptimal methods, validating its effectiveness in balancing domain adaptation, continuous segmentation, and knowledge retention under limited annotations, thereby providing a feasible solution for continuous learning in dynamic clinical data environments.

摘要： 为了解决医学图像分割场景下半监督域增量学习面临的标注数据稀缺、累积分布差异及域与类别层面特征漂移导致灾难性遗忘的问题，需探索能在有限标注条件下持续适应新域并保留已有分割能力的有效方法。针对上述问题，提出了一种双粒度原型对比学习方法，首先利用最优传输驱动半监督特征对齐以协调分布差异，其次通过原型对比学习在域增量学习过程中动态适应特征分布偏移，最后引入历史原型推理校正机制利用历史原型信息优化模型的预测结果，有效解决了稀疏标注资源、动态数据分布与分割目标在增量学习中的冲突。基于心脏与前列腺分割数据集的实验结果表明，在稀疏标注下，该方法展现出对新域的持续适应能力、稳定的连续分割性能及一定的抗遗忘性。实验结果表明，相比基线方法，所提方法在多个增量阶段中的总体Dice相似系数在心脏和前列腺数据集上分别达到了87.13%，86.05%，与次优方法对比效果分别提升了5.98和18.64个百分点，验证了在有限标注下兼顾域适应、连续分割与知识保持的有效性，为临床动态数据环境中的持续学习提供了可行思路。