Paper Reading
BTS-DSN: Deeply supervised neural network with short connections forretinal vessel segmentation
references
: you can read this papar via BTS-DSN or star/fork this github project
Abstract
- An important method was proposed by this group called short-connection, that improve quite a lot the result of this model’s ability. They use sensitivity,specificity, AUC and F1-score to test their model. If you aren’t familiar with these indicators, you can read these two blogs to learn moreAUC &AUC-2.
1. Introduction
- why we hope to achieve early dignosis.
- some researchs done by other scholar group
(1) Unsupervised methods: some methods have been proposed by scholoars, mainly about detect the profile and contour of vessel. These methods are mostly based on geometric computation model. defects:
However, the unsupervised methods are sensitive to the manually designed features and rules.
(poor in generalization)
(2)supervised methods:when we use supervised method, we usually view this problem as a pixel-wise binary classification. deep learning methods are popular in this area. Other ways to solve this problem about semantic segmentation results.
defects:different methods have different disadvantages such as time-consuming and so on
- the contributions of this paper
(1)”We propose a deeply-supervised fully convolutional neural network with bottom-top and top-bottom short connections (BTS-DSN) for vessel segmentation. “
(2)”We used VGGNet and ResNet-101 as backbone and conducted extensive experiments on DRIVE, STARE and CHASE_DB1”
“We employed cross-training experiments to show the generalization of BTS-DSN.”
attributes : more about VGGNet and ResNet can be found in their papers.{vgg(including vgg-16&vgg-19)&ResNet_paper&ResNet_github}.If you want to understand these model better, baidu or google it may help you.
2.BTS-DSN
from HED to DSN, and then to BS-DSN and BST-DSN. this method was proposed to alleviate the gradient vanish problem in deep network. It’s a deep supervision
bottom-top short connnections: pass low level fine semantic information to high levels to alleviate the blurring situation.
top-bottom short connection: Bottom-top short connections aim to refine high-level segmentation results.
inference: do feature confusion
3. Implementation details
data augmentation: using quite a lot various transformations to augment the training set, including rotation, flipping and scaling.
model implementation : using the network framework Caffe, short connections of the BTS-DSN.
Parameter settings
When the backbone is VGGNet, we fine-tuned our network with a
learning rate of 1e-8, a weight decay of 0.0005, and a momentum of
0.9. We use a fixed learning rate.
in two different backbone(VGGNet & ResNet-101),they use different type of parameters .
and for patch-level S-DSN, they split a raw retinal image into 9 patches, each of which was 1/4 the size of the raw image, meanwhile, patches were up-sampled $2 \times$
4.running environment:……
Evaluation criteria
In vessel segmentation, each pixel belongs to a vessel or non-vessel
pixel.
(be seen as a binary classfication problem).they iemployed six evaluation criteria, including AUC,SE,SP,ACC,F1-score and MCC.
Result
the result are shown in the fig in this paper, so you can just look through and get them.
Conclusion
这篇文章成功实现了目前基本上是最高水平的眼底血管分类成果,其使用VGGNet或者ResNet-101作为支柱(backbone)来实现整个CNN卷积的过程,同时,为了缓解语义分割中语义分割之间的差距,使用top-bottom 和 bottom-top short connection来实现整个过程,最终实现了BTS-DSN网络,达到了好的效果。同时,这个网络也使用了特征融合的方法,可以加以了解。对于更多的细节,还需要继续了解。