中文解析可泛化超声基础模型的任务聚合

该研究指出，超声基础模型在多任务统一时性能下降的主因并非容量不足，而是任务聚合策略忽略了异质性与数据规模的交互作用。通过系统分析，建立了异构超声任务无性能损失联合学习的实用标准，为多任务模型设计提供指导。

arXiv:2603.18123v3 Announce Type: replace Abstract: Foundation models promise to unify multiple clinical tasks within a single framework, but recent ultrasound studies report that unified models can underperform task-specific baselines. We hypothesize that this degradation arises not from model capacity limitations, but from task aggregation strategies that ignore interactions between task heterogeneity and available training data scale. In this work, we systematically analyze when heterogeneous ultrasound tasks can be jointly learned without performance loss, establishing practical criteria f