Hu Zhaoyu, Sun Yuhao, Bian Liuguan, Luo Chun, Zhu Junle, Zhu Jin, Li Shiting, Zhao Zheng, Wang Yuanyuan, Shi Huidong, Shi Zhifeng, Yu Jinhua
School of Information Science and Technology, Fudan University, Shanghai, China.
Department of Neurosurgery, Ruijin Hospital, Shanghai Jiao Tong University, Shanghai, China.
Comput Biol Med. 2025 Feb;185:109472. doi: 10.1016/j.compbiomed.2024.109472. Epub 2024 Dec 4.
Gliomas are the most common and malignant form of primary brain tumors. Accurate segmentation and measurement from MRI are crucial for diagnosis and treatment. Due to the infiltrative growth pattern of gliomas, their labeling is very difficult. In turn, the already available annotated datasets, such as well-known BraTS, are difficult to generalize to multicenter unannotated datasets due to the variations in imaging machines and parameters. To address this challenge, a novel unsupervised domain adaptation framework for glioma segmentation (UDA-GS) is proposed. UDA-GS uses GliomaMix to mix labeled tumors with unlabeled images and aligns the features of the same tumor, allowing the network to adapt to different backgrounds across different centers. Additionally, the framework leverages tumor information generated by GliomaMix as prior knowledge for self-supervised regression tasks to enhance feature encoding for tumors in different domains. Using Mean-Teacher as the basic framework, UDA-GS also incorporates weighted consistency regularization and mask combining strategy to achieve efficient unsupervised domain adaptation. Quantitative and qualitative evaluations were conducted on 1179 cases across 27 centers, without requiring any local annotations. The results demonstrate that UDA-GS outperforms the second-best method in terms of Dice coefficient segmentation metrics by 18.2 %, 6.9 %, and 4.6 % for the whole tumor, tumor core, and enhanced tumor, respectively, on the internal testing set. Additionally, the evaluations reveal that doctors express greater satisfaction with the segmentation outcomes achieved by UDA-GS in comparison to other methods including the segment anything model (SAM).
胶质瘤是原发性脑肿瘤中最常见且恶性程度最高的形式。通过磁共振成像（MRI）进行准确的分割和测量对于诊断和治疗至关重要。由于胶质瘤的浸润性生长模式，对其进行标注非常困难。相应地，现有的带注释数据集，如著名的脑肿瘤多模态数据集（BraTS），由于成像设备和参数的差异，难以推广到多中心无注释数据集。为应对这一挑战，提出了一种用于胶质瘤分割的新型无监督域适应框架（UDA - GS）。UDA - GS使用GliomaMix将带标签的肿瘤与无标签图像混合，并对齐同一肿瘤的特征，使网络能够适应不同中心的不同背景。此外，该框架利用GliomaMix生成的肿瘤信息作为自监督回归任务的先验知识，以增强不同域中肿瘤的特征编码。以均值教师（Mean - Teacher）为基本框架，UDA - GS还结合了加权一致性正则化和掩码合并策略，以实现高效的无监督域适应。在27个中心的1179例病例上进行了定量和定性评估，无需任何局部注释。结果表明，在内部测试集上，就全肿瘤、肿瘤核心和强化肿瘤的骰子系数分割指标而言，UDA - GS分别比次优方法高出18.2%、6.9%和4.6%。此外，评估还显示，与包括分割一切模型（SAM）在内的其他方法相比，医生对UDA - GS实现的分割结果满意度更高。
