3D U-Net笔记

一、Network Architecture

Like the standard u-net, it has an analysis and a synthesis path.

In the analysis path, 3 × 3 × 3 convolutions, 2 × 2 × 2 max pooling with strides of two.

In the synthesis path, upconvolution of 2×2×2 by strides of two, 3×3×3 convolutions.

Doubling the channel before max pooling.

Introduce BN before each ReLU.

二、Two cases of this method

Semi-automated: the user annotates some slices of each volume to be segmented. The network predicts the dense segmentation.

Fully-automated: the network is trained with annotated slices from a representative training set and can be run on non-annotated volumes

2.1Semi-automated Segmentation

The network output and the ground truth labels are compared using softmax with weighted cross-entropy loss, where we reduce weights for the frequently seen background and increase weights for the inner tubule to reach a balanced influence of tubule and background voxels on the loss. Weighted softmax loss function allows us to train on sparse annotations. Setting the weights of unlabeled pixels to zero makes it possible to learn from only the labelled ones.

（在训练semi-automatic时，一个立方体中只有一些slices是被标注的，包括xy，xz，yz三个平面。在最后的loss中，不考虑未被标注的slices中的pixel，也就是weights置0。 根据已知label的pixel训练网络，然后用在同个立方体中的其他未知label的pixel上。）

2.2Fully-automated Segmentation

将semi-automated的结果用在完全无标注的立方体中。