Deep Learning Techniques for Medical Image Segmentation: Achievements and Challenges

阅读：评论：0

前言：
该篇文章较为全面但稍偏简单的介绍医学图像分割的常见数据集、各种神经网络，以及常见的训练技巧等问题。
一、重点摘录

2.5D approaches are inspired by the fact that 2.5D has the richer spatial information of neighboing pixels wiht less computational costs than 3D.
The 3D network is trained to predict the label of a central voxel according to the content fo surrounding 3D patches.
To address the deimensionality issue and reduce the processing time, Dou et al. in [23] proposed to utilize a set of 3D kernels that shared the weights spatially, which helped to reduce the number of parameters.
The authors in [66] applied a hierarchical coarse-to-fine strategy that significantly improved the segmentation results of small organs.
Focal FCN
Zhou et al. [93] proposed to apply the focal loss on the FCN to reduce the number of false positives occurred due to the unbalanced ration of background and foreground pixels in medical images. In this structure, the FCN was used to produce the intermediate segmentation results and then the focal FCN was used to remove the false positives.
In an attempt to enpower the U-Net structure with richer spatial information, Cicek et al. developed a 3D U-Net model [17]. The average IoU (i.e., Intersection over Union) of 0.863 demonstrated that network was able to find the whole 3D volume from few annotated slices successfully by using a weighted softmax loss function.
The deeper networks not only suffer from gradient vanishing problem but also face the more pressing issue of degradation [33]. It means with the depth increasing, the accuracy gets saturated and then rapidly degrades. To take advantage from deeper network structure, He et al. [33] introduced the residual networks which were initially developed for natural image segmentation on 2D images.
Data Augmentation
references papers: [30] [52]
Class Imbalance
A popular solution for this issue is sample re-weighting, where a higher weight is applied to the foreground patches during training [16]. Automatic modification of sample re-weighting has been developed by using Dice loss layer and Dice coefficent [44,62,95]. Yet, the effectiveness is limited in dealing with extreme class imbalance [93]. Patch-wise training combind with patch selection can help to address the issue of class imbalance [18]. Fundamentally, during the creation of the training set, a control mechanism can be set to have a balanced number of pathces from the background and foreground [52]. Another approach to deal with this issue is sampled loss in which the loss will not be calculated for the entire image and just some random pixels(areas) will be selected for loss calculation [56].