📄 中文摘要
医疗图像分割的性能越来越取决于数据利用效率，而非仅仅原始数据量。特别是对于脑膜瘤等复杂病理的精确分割，模型需要充分利用有限高质量标注中的潜在信息。为最大限度地发挥现有数据集的价值，本研究提出了一种新颖的双重增强框架，协同整合了两种关键策略。首先，引入隐式时空混合（Implicit Spatiotemporal Mixing, ISM）模块，通过在深度特征空间中动态融合来自不同时间点或空间位置的信息，增强了模型对脑膜瘤形态和演变模式的理解。ISM通过学习非线性映射，将多模态或多时相数据中的互补信息进行有效整合，从而在数据量有限的情况下提升了特征的判别力。其次，开发了模拟到真实（Sim2Real）语义注入机制，旨在弥合合成数据与真实世界数据之间的领域差距。通过利用大量可生成的合成脑膜瘤图像，并结合对抗训练或领域适应技术，将合成数据中学到的丰富语义知识有效地转移到真实数据的分割任务中。Sim2Real语义注入通过生成对抗网络（GANs）或其他生成模型，创建具有多样化外观和病理特征的合成脑膜瘤图像，随后通过特定的训练范式，如循环一致GAN或特征空间对齐，将这些合成图像的知识迁移至真实世界数据，从而极大地扩充了训练数据并增强了模型的泛化能力。该框架通过结合ISM对数据内在信息的深度挖掘和Sim2Real对外部知识的有效引入，显著提升了在数据稀缺场景下脑膜瘤分割的准确性和鲁棒性，尤其适用于临床上标注成本高昂、数据获取困难的挑战。
📄 English Summary
Data-Efficient Meningioma Segmentation via Implicit Spatiotemporal Mixing and Sim2Real Semantic Injection
The efficacy of medical image segmentation is increasingly dictated by the efficiency of data utilization rather than the sheer volume of raw data. Accurate segmentation, particularly for intricate pathologies like meningiomas, necessitates that models comprehensively leverage the latent information embedded within limited high-quality annotations. To maximize the utility of existing datasets, a novel dual-augmentation framework is proposed, synergistically integrating two pivotal strategies. Firstly, an Implicit Spatiotemporal Mixing (ISM) module is introduced, designed to enhance the model's understanding of meningioma morphology and evolutionary patterns by dynamically fusing information from disparate temporal points or spatial locations within the deep feature space. ISM achieves this by learning non-linear mappings to effectively integrate complementary information across multi-modal or multi-temporal data, thereby improving feature discriminability even with limited data. Secondly, a Sim2Real Semantic Injection mechanism is developed to bridge the domain gap between synthetic and real-world data. This mechanism exploits a large volume of generatable synthetic meningioma images and, through adversarial training or domain adaptation techniques, effectively transfers the rich semantic knowledge acquired from synthetic data to the segmentation task involving real data. Sim2Real semantic injection employs generative adversarial networks (GANs) or other generative models to create synthetic meningioma images with diverse appearances and pathological characteristics. Subsequently, specific training paradigms, such as cycle-consistent GANs or feature space alignment, are utilized to transfer the knowledge from these synthetic images to real-world data, thereby significantly augmenting the training data and enhancing the model's generalization capabilities. This framework, by combining ISM's deep exploitation of intrinsic data information and Sim2Real's effective incorporation of external knowledge, substantially improves the accuracy and robustness of meningioma segmentation in data-scarce scenarios, making it particularly suitable for clinical challenges where annotation costs are high and data acquisition is difficult.