Abstract:

Segmented medical images for diagnosis serve as an effective auxiliary method in clinical and medical research. However, challenges such as privacy concerns, data dispersion, and labeling difficulties hinder their practical application. For example, in the case of cervical spine Magnetic Resonance Imaging (MRI) image segmentation, it is more challenging to obtain image data, and labeling costs are high. It is difficult for the cervical spine segmentation model to effectively extract detailed cervical spine information when encountering heterogeneous data from different sources. Therefore, in the federated learning scenario, this paper proposes M-FedLO. It is a multi-scale semi-supervised segmentation network based on label separation and bootstrapping. It aims to address the challenges of limited labeled information and reduced accuracy when handling heterogeneous data. M-FedLO employs label separation to separately segment the vertebrae and intervertebral discs while achieving multi-scale outputs. This allows for further edge information extraction and better separation between vertebral blocks and intervertebral discs. In the ″global + local″ mode of federated learning, the labels from the global model guide the local models. This ensures that the features extracted by the local models from unlabeled data approximate those of the global model. This, in turn, enhances the utilization of unlabeled data by the local models. Additionally, the approach uses Stochastic Weight Averaging(SWA) to optimize the parameters to alleviate model weight oscillation issues and enhance the model's generalization capability. The experimental results demonstrate that the proposed model outperforms the semi-supervised baseline segmentation models in the segmentation of non-heterogeneous cervical spine MRI medical images and heterogeneous cervical spine images. Specifically, compared with the best-performing ICT model on the cervical spine dataset, the proposed approach achieves an improvement in the Dice Similarity Coefficient(DSC) metric to 86.86%. This represents an enhancement by 1.72 percentage points.

Key words: cervical spine segmentation, federated learning, heterogeneous data, label separation, multi-scale, label-guided

摘要：

利用分割的医学图像进行诊断在临床和医学研究上是一种有效的辅助方法, 但由于医学图像的隐私性、分散性和标注困难等问题严重影响了其实际应用效果。对颈椎磁共振成像(MRI)图像分割来说, 其图像数据获取更困难, 且标注成本高昂, 颈椎分割模型在面对不同来源的异质性数据时难以有效提取颈椎细节信息。因此, 在联邦学习场景下, 针对标注信息缺少以及数据异质性导致分割精度下降的问题, 提出一种基于标签分离与引导的多尺度半监督分割网络M-FedLO。M-FedLO通过标签分离的方式分别对椎块与椎间盘进行分割, 同时实现多尺度输出, 使得椎块与椎间盘的边缘信息得到进一步提取, 更好地分离出椎块与椎间盘。在联邦“全局+本地”的模式下, 利用全局模型的标签引导, 使本地模型在无标签数据上提取的特征与全局模型逼近一致, 从而增强本地模型对无标签数据的利用。同时使用随机权重平均(SWA)算法对参数进行优化, 缓解模型权重震荡问题, 提升模型泛化能力。实验结果表明, 与半监督基准分割模型相比, 提出的模型不仅在非异质性上的颈椎MRI医学图像分割效果上取得一定的提升, 而且在异质性的颈椎图像上也具有较好的成果。在颈椎数据集上与实验结果最好的ICT模型相比较, Disc相似性系数(DSC)指标达到86.86%, 提升了1.72个百分点。

关键词: 颈椎分割, 联邦学习, 异质性数据, 标签分离, 多尺度, 标签引导