【摘要】 生物医学图像配准与分割在生命科学研究、医学诊断与临床治疗等领域具有重要意义，然而其深度学习方法通常依赖于大规模高质量标注数据，获取成本高昂。现有数据增广技术多基于旋转、平移等线性变换方法生成额外数据，仍难以有效解决生物医学图像对数据亮度多样性与柔性结构形变多样性的需求问题。对此，提出了一种基于配准和分割联合模型的双采样数据增广策略。通过结合亮度场和形变场上的双重采样提升了生成数据的亮度多样性和柔性结构形变多样性。此外，为抑制伪标注数据中错误信息对联合模型产生的误导，采用对抗式训练方式设计了伪标注鉴别器模块来寻找错误的分割预测，抑制了错误信息在数据循环过程中的传播。最后，设计了一种单样本场景下的配准-分割协同框架，在Mindboggle-101人脑MRI数据集、MouseBrain鼠脑fMOST数据集和MM-WHS 2017心脏数据集上，仅使用一幅带标签的训练图像实现了优于对比算法的分割和配准性能，验证了所提数据增广策略的有效性。更多还原

【Abstract】 Biomedical image registration and segmentation are essential in life science research, medical diagnosis, and clinical treatment. However, deep learning methods typically rely on large-scale, high-quality annotated data, which are costly to obtain. The existing data augmentation techniques mainly generate additional samples through linear deformation such as rotation and translation methods, yet they still struggle to effectively solve the demand for data brightness diversity and elastic structure deformation diversity of biomedical images.In this paper, we propose a double-sampling data augmentation strategy based on the joint model of registration and segmentation. Integrating the double sampling on the brightness and deformation fields enriches the brightness diversity and the data’s elastic structural deformation diversity.Further, to suppress the misleading of incorrect information in the pseudo-labeled data on the joint model, an adversarial training method is used to train a pseudo-label discriminator module to find inaccurate segmentation predictions and suppress the propagation of incorrect information in the data cycle. Eventually, this paper proposes a joint model for registration and segmentation for a one-shot scenario. It achieves better segmentation and registration performance than the comparison algorithms on the Mindboggle-101 human brain MRI dataset, MouseBrain mouse brain fMOST dataset, and MM-WHS 2017 heart dataset, demonstrating the effectiveness of our data augmentation strategy.更多还原