Department of Robotics and Mechatronics, The Faculty of Electrical Engineering, Mathematics and Computer Science, Technical Medical Centre, University of Twente, 7522 NB, Enschede, The Netherlands.
Int J Comput Assist Radiol Surg. 2021 Jan;16(1):41-51. doi: 10.1007/s11548-020-02292-y. Epub 2020 Nov 21.
We investigated the parameter configuration in the automatic liver and tumor segmentation using a convolutional neural network based on 2.5D model. The implementation of 2.5D model shows promising results since it allows the network to have a deeper and wider network architecture while still accommodates the 3D information. However, there has been no detailed investigation of the parameter configurations on this type of network model.
Some parameters, such as the number of stacked layers, image contrast, and the number of network layers, were studied and implemented on neural networks based on 2.5D model. Networks are trained and tested by utilizing the dataset from liver and tumor segmentation challenge (LiTS). The network performance was further evaluated by comparing the network segmentation with manual segmentation from nine technical physicians and an experienced radiologist.
Slice arrangement testing shows that multiple stacked layers have better performance than a single-layer network. However, the dice scores start decreasing when the number of stacked layers is more than three layers. Adding higher number of layers would cause overfitting on the training set. In contrast enhancement test, implementing contrast enhancement method did not show a statistically significant different to the network performance. While in the network layer test, adding more layers to the network architecture does not always correspond to the increasing dice score result of the network.
This paper compares the performance of the network based on 2.5D model using different parameter configurations. The result obtained shows the effect of each parameter and allow the selection of the best configuration in order to improve the network performance in the application of automatic liver and tumor segmentation.
我们研究了基于 2.5D 模型的卷积神经网络自动肝脏和肿瘤分割的参数配置。2.5D 模型的实现显示出了有前景的结果，因为它允许网络具有更深和更宽的网络架构，同时仍然适应 3D 信息。然而，对于这种类型的网络模型，还没有对参数配置进行详细的研究。
在基于 2.5D 模型的神经网络上研究和实现了一些参数，例如堆叠层数、图像对比度和网络层数。利用肝脏和肿瘤分割挑战赛（LiTS）的数据集对网络进行训练和测试。通过比较网络分割与来自九位技术医生和一位经验丰富的放射科医生的手动分割，进一步评估了网络性能。
切片排列测试表明，多层堆叠比单层网络具有更好的性能。然而，当堆叠层数超过三层时，骰子得分开始下降。增加更多的层数会导致训练集过拟合。相反，在对比度增强测试中，实施对比度增强方法对网络性能没有统计学上的显著差异。而在网络层测试中，向网络架构中添加更多的层并不总是对应于网络骰子得分的增加。
本文比较了基于不同参数配置的 2.5D 模型网络的性能。所得结果显示了每个参数的效果，并允许选择最佳配置，以提高自动肝脏和肿瘤分割应用中的网络性能。
