RM Transunet：基于小样本数据的肺癌脑转移瘤分割

RM Transunet: Segmentation of Lung Cancer Brain Metastases Based on Small Sample Data

摘要: 脑肿瘤的语义分割是一项基本的医学图像分析任务，可以协助临床医生诊断患者，并持续关注病灶部分的变化情况。得益于深度学习的发展，医学图像自动分割取得了很大的进步。然而，现有的深度学习分割模型依赖于庞大的训练数据支撑，在临床上，医学数据通常数据量较小。为了改善这些问题，本文提出了一种新的深度医学图像分割框架，称为Residual Mulitse Transunet (RM Transunet)，以提高小样本医学图像的语义分割质量。本文提出的RM Transunet遵循了Transunet的设计，融合了CNN和Transformer，并引入了Residual Block和MuilSE，有效地在不同尺度的特征之间建立全局依赖关系，以便充分利用这些获得的多尺度特征表示。针对医学图像分割的实验证明了RM Transunet的有效性，并表明我们的方法明显优于当下的方法。本研究的贡献不仅在于提供了一种新的思路来解决医学图像领域小样本数据的挑战，同时也为临床决策和治疗优化提供了有益的参考。

Abstract: Semantic segmentation of brain tumors is a basic medical image analysis task that can assist clinicians in diagnosing patients and continuously paying attention to changes in the lesions. Thanks to the development of deep learning, great progress has been made in automatic segmentation of medical images. However, existing deep learning segmentation models rely on the support of huge training data. In clinical practice, medical data usually have a small amount of data. To improve these problems, this paper proposes a new deep medical image segmentation framework called residual mulitse transunet (RM Transunet) to improve the semantic segmentation quality of small sample medical images. The RM Transunet proposed in this article follows the design of Transunet, integrates CNN and Transformer, and introduces Residual Block and MuilSE to effectively establish global dependencies between features at different scales in order to make full use of these obtained multi-scale feature representations. Experiments on medical image segmentation demonstrate the effectiveness of RM Transunet and show that our method significantly outperforms current methods. The contribution of this study is not only to provide a new idea to solve the challenge of small sample data in the field of medical images, but also to provide a useful reference for clinical decision-making and treatment optimization.