医学图像分割
Topology Aware Fully Convolutional Networks For Histology Gland Segmentation
Lecture Notes in Computer Science, Medical Image Computing and Computer-Assisted Intervention (MICCAI), volume 9900, pages 460-468, 2016
code: http://www.sfu.ca/~abentaie/topo_fcn/topo_fcn.html
本文将 FCN 加入 geometric and topological 先验信息 用于 组织细胞学图像分割
首先来看看 FCN 的缺点吧
FCN-based segmentations suffer from relying on a pixel-level prediction that is not designed to account for higher-order properties, such as boundary smoothness and the topological label interactions of multi-part objects (as in the lumen and epithelium of glands). FCN 针对像素级别的分割，没有考虑图像的higher-order 属性例如 边缘平滑性，多物体之间的拓扑信息
Moreover, FCNs tend to produce low-resolution segmentations due to the subsampling resulting from stacked layers of convolutions and pooling. 由于降采样的原因导致 FCN 的输出结果分辨率低，比较粗糙
虽然已经有人通过 引入 conditional random field (CRF) 来改善 FCN，但是计算量比较大，同时只有特定的图模型被嵌入到 FCN中 only specific graphical models can be integrated into the FCN learning pipeline。 这里我们主要是将 topology priors 嵌入到 FCN中
主要是通过定义一个特定的损失函数来实现的 Our strategy is to design a loss function with specific penalty terms that encode the desired boundary smoothness priors and hierarchical relationships between regions labels
Multi-Part Object Segmentation
Proposed Topology-Augmented Loss
Experimental Performance Evaluation
