UNet-family
U-Net
[paper] : U-Net: Convolutional Networks for Biomedical Image Segmentation
根据 UNet的结构，它能够结合底层和高层的信息。
底层(深层)信息：经过多次下采样后的低分辨率信息。能够提供分割目标在整个图像中上下文语义信息，可理解为反应目标和它的环境之间关系的特征。这个特征有助于物体的类别判断（所以分类问题通常只需要低分辨率/深层信息，不涉及多尺度融合）
高层(浅层)信息：经过concatenate操作从encoder直接传递到同高度decoder上的高分辨率信息。能够为分割提供更加精细的特征，如梯度等。
V-Net
[paper]:V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation
相当于3D的unet
H-DenseUNet
[paper]: H-DenseUNet: Hybrid Densely Connected UNet for Liver and Tumor Segmentation from CT Volumes
网络可以很好的提取混合特征：片内特征和片间的融合特征
网络很好的解决了用2D训练缺少体积上下文信息以及3D训练需要消耗高额计算内存的问题
每一个3D输入，通过变换处理函数F，将3D的体积快变成2D的相邻的slices；然后将这些2D slices送进2D DenseUNet，用来提取片内特征；3D的原始输入和2D DenseUNet(用于有效的片内2D特征提取)转换之后的预测结果concat在一起送进3D网络3D DenseUNet(用于上下文信息探测)，用来提取片间特征；然后将两者特征融合并经由HFF层(用于联合优化片内特征和片间特征的混合特征融合层)预测最终结果。
原始3D volume上下各填充一层slice 按照 012、 123、 234、 345 … 的方式，将3D volume变换成 2D slices。
UNet++
[paper]: UNet++: A Nested U-Net Architecture for Medical Image Segmentation
https://zhuanlan.zhihu.com/p/44958351
MDU-Net
[paper]: MDU-Net: Multi-scale Densely Connected U-Net for biomedical image segmentation
DANet
[paper]:
DUNet
[paper]: DUNet: A deformable network for retinal vessel segmentation
Deformable U-Net
Deformable Convolution
Structure
Multi-path U-Net
[paper]: Dense Multi-path U-Net for Ischemic Stroke Lesion Segmentation in Multiple Image Modalities
Bridged U-net
[paper]: Prostate Segmentation using 2D Bridged U-net
SUNet
[paper]: SUNet: a deep learning architecture for acute stroke lesion segmentation and outcome prediction in multimodal MRI
LadderNet
LADDERNET: Multi-Path Networks Based on U-Net for Medical Image Segmentation [paper]
Attention U-Net
[paper]: Attention U-Net: Learning Where to Look for the Pancreas
M-Net
[paper]: M-NET: A Convolutional Neural Network for Deep Brain Structure Segmentation
Improved Attention M-Net
[paper]: A Novel Focal Tversky Loss Function with Improved Attention U-Net for Lesion Segmentation
R2U-Net
[paper]: Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net) for Medical Image Segmentation
SE-Net
[paper]: Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks
Y-Net
[paper]: Y-Net: Joint Segmentation and Classification for Diagnosis of Breast Biopsy Images
(input: 2D)
U-Net vs Y-Net
[paper]: Detection and Delineation of Acute Cerebral Infarct on DWI Using Weakly Supervised Machine Learning (Y-Net)
(input: 3D)
MultiResUNet
[paper]: MultiResUNet : Rethinking the U-Net Architecture for Multimodal Biomedical Image Segmentation
CE-Net
[paper]: CE-Net: Context Encoder Network for 2D Medical Image Segmentation
Graph U-Net
[paper]: Graph U-Net
CSA U-NET
[paper]: Connection Sensitive Attention U-NET for Accurate Retinal Vessel Segmentation
CIA-Net
[paper]: CIA-Net: Robust Nuclei Instance Segmentation with Contour-aware Information Aggregation
W-Net
[paper]: W-Net: A Deep Model for Fully Unsupervised Image Segmentation
[paper]: W-Net: Reinforced U-Net for Density Map Estimation
ScleraSegNet
[paper]: ScleraSegNet: an Improved U-Net Model with Attention for Accurate Sclera Segmentation
AHCNet
[paper]: AHCNet: An Application of Attention Mechanism and Hybrid Connection for Liver Tumor Segmentation in CT Volumes
Probabilistic U-Net
[paper]: A Hierarchical Probabilistic U-Net for Modeling Multi-Scale Ambiguities
Partially Reversible U-Net
[paper]: A Partially Reversible U-Net for Memory-Efficient Volumetric Image Segmentation
RA-UNet
[paper]: RA-UNet: Residual Attention U-Net for Semantic Segmentation of Cataract Surgical Instruments
Unet-GAN
[paper]: The Domain Shift Problem of Medical Image Segmentation and Vendor-Adaptation by Unet-GAN
Siamese U-Net
[paper]:Siamese U-Net with Healthy Template for Accurate Segmentation of Intracranial Hemorrhage
U-Det:
Losses
https://blog.csdn.net/m0_37477175/article/details/83004746
Cross Entropy Loss
- binary cross entropy 二分类交叉熵
- categorical cross entropy 多分类交叉熵
- weighted cross entropy 加权交叉熵
缺点：交叉熵的损失函数单独评估每个像素矢量的类预测，然后对所有像素求平均值。但是图像中常出现类别不均衡（class imbalance）的问题，由此导致训练会被像素较多的类主导，对于较小的物体很难学习到其特征，从而降低网络的有效性
Dice Coefficient Loss
\[\begin{equation} \text {Dice}=\frac{2|A \cap B|}{|A\|+|B|} \end{equation}\]该指标范围从0到1，其中1表示完整的重叠，loss=1-dice
dice loss比较适用于样本极度不均的情况
IOU Loss
\[I O U=\frac{|A \cap B|}{|A|+|B|-|A \cap B|}\]loss=1-IOU
Focal Loss
可以改善目标不均衡的现象
Generalized Dice loss
在使用DICE loss时，对小目标是十分不利的，因为在只有前景和背景的情况下，小目标一旦有部分像素预测错误，那么就会导致Dice大幅度的变动，从而导致梯度变化剧烈，训练不稳定。
the generalized Dice loss在dice loss基础上给每个类别加权，起到平衡各类（包括背景类）目标区域对loss的贡献。
Tversky loss
\[T(A, B)=\frac{|A \cap B|}{|A \cap B|+\alpha|A-B|+\beta|B-A|}\]是dice系数和jaccard系数的一种广义系数
