Kuang Xihe, Cheung Jason Py, Wu Honghan, Dokos Socrates, Zhang Teng
Annu Int Conf IEEE Eng Med Biol Soc. 2020 Jul;2020:1633-1636. doi: 10.1109/EMBC44109.2020.9175987.
Most deep learning based vertebral segmentation methods require laborious manual labelling tasks. We aim to establish an unsupervised deep learning pipeline for vertebral segmentation of MR images. We integrate the sub-optimal segmentation results produced by a rule-based method with a unique voting mechanism to provide supervision in the training process for the deep learning model. Preliminary validation shows a high segmentation accuracy achieved by our method without relying on any manual labelling.The clinical relevance of this study is that it provides an efficient vertebral segmentation method with high accuracy. Potential applications are in automated pathology detection and vertebral 3D reconstructions for biomechanical simulations and 3D printing, facilitating clinical decision making, surgical planning and tissue engineering.
大多数基于深度学习的椎体分割方法都需要繁重的手动标注任务。我们旨在建立一种用于磁共振图像椎体分割的无监督深度学习流程。我们将基于规则的方法产生的次优分割结果与独特的投票机制相结合，以便在深度学习模型的训练过程中提供监督。初步验证表明，我们的方法在不依赖任何手动标注的情况下实现了很高的分割精度。这项研究的临床意义在于它提供了一种高效且高精度的椎体分割方法。潜在应用包括自动病理检测以及用于生物力学模拟和3D打印的椎体三维重建，有助于临床决策、手术规划和组织工程。
