Comparing Deep Learning Strategies and Attention Mechanisms of Discrete Registration for Multimodal Image-Guided Interventions
Comparing Deep Learning Strategies and Attention Mechanisms of Discrete Registration for Multimodal Image-Guided Interventions
复制标题
DOI：
10.1007/978-3-030-33642-4_16
复制
发表时间：
2019-10
期刊：
影响因子：
--
通讯作者：
I. Ha;M. Heinrich
中科院分区：
文献类型：
--
作者：
I. Ha;M. Heinrich
在医学成像中，深度学习已成功应用于分割和分类任务，但其在图像配准任务中的应用仍然有限。离散配准的使用可以通过帮助捕获更复杂的变形来缓解限制基于CNN的配准用于大位移的问题。我们评估了不同的构建块的学习为基础的离散配准的CuRIOUS多模态图像配准的挑战。我们还提出了一个新的注意力模块，估计网格点的信息内容，比较不同的损失函数和评估的自我监督的预训练的特征提取步骤的影响。
In medical imaging, deep learning has been applied to segmentation and classification tasks successfully, whereas its use for image registration tasks is still limited. The use of discrete registration can alleviate the problems limiting the use of CNN based registration for large displacements by helping to capture more complex deformations. We evaluate different building blocks of learning based discrete registration for the CuRIOUS multimodal image registration challenge. We also propose a new attention module, which estimates information contents of a grid point, compare different loss functions and evaluate the influence of self-supervised pre-training of feature extraction step.