Central South University, No. 932 Lushan South Road, Changsha, 410000, China.
Xiangya Hospital of Central South University, No. 87 Xiangya Road, 410000, China.
Comput Biol Med. 2023 Jun;160:106908. doi: 10.1016/j.compbiomed.2023.106908. Epub 2023 Apr 24.
Accurate tissue segmentation on MRI is important for physicians to make diagnosis and treatment for patients. However, most of the models are only designed for single-task tissue segmentation, and tend to lack generality to other MRI tissue segmentation tasks. Not only that, the acquisition of labels is time-consuming and laborious, which remains a challenge to be solved. In this study, we propose the universal Fusion-Guided Dual-View Consistency Training(FDCT) for semi-supervised tissue segmentation on MRI. It can obtain accurate and robust tissue segmentation for multiple tasks, and alleviates the problem of insufficient labeled data. Especially, for building bidirectional consistency, we feed dual-view images into a single-encoder dual-decoder structure to obtain view-level predictions, then put them into a fusion module to generate image-level pseudo-label. Moreover, to improve boundary segmentation quality, we propose the Soft-label Boundary Optimization Module(SBOM). We have conducted extensive experiments on three MRI datasets to evaluate the effectiveness of our method. Experimental results demonstrate that our method outperforms the state-of-the-art semi-supervised medical image segmentation methods.
MRI 上的精确组织分割对于医生为患者进行诊断和治疗非常重要。然而，大多数模型仅针对单任务组织分割进行设计，往往缺乏对其他 MRI 组织分割任务的通用性。不仅如此，标签的获取既费时又费力，这仍然是一个有待解决的挑战。在这项研究中，我们提出了用于 MRI 上半监督组织分割的通用融合引导双视图一致性训练（FDCT）。它可以为多个任务获得准确和鲁棒的组织分割，并缓解标记数据不足的问题。特别是，为了构建双向一致性，我们将双视图图像输入到单个编码器双解码器结构中以获得视图级别的预测，然后将其放入融合模块中以生成图像级别的伪标签。此外，为了提高边界分割质量，我们提出了软标签边界优化模块（SBOM）。我们在三个 MRI 数据集上进行了广泛的实验，以评估我们方法的有效性。实验结果表明，我们的方法优于最先进的半监督医学图像分割方法。
