Retinal Vessel Segmentation via a Semantics and Multi-Scale Aggregation Network

subject ： Retinal vessel segmentation based on semantic and multi-scale aggregation network    2020

author ： Xu Rui     School of software, Dalian University of technology

Data sets ：CHASE_DB1 and HRF

meeting ：ICASSP 2020-2020 IEEE acoustics 、 International Conference on speech and signal processing （ICASSP）

Purpose ： Improve the segmentation accuracy of capillaries .

Method ：
 First, the higher-order semantic information with stronger representation ability is gathered through the semantic aggregation module , And then get a strong feature expression , To accurately segment capillaries and blood vessel junctions .
 In addition, a multi-scale aggregation module is designed , Parallel branches with different expansion rates are used to extract multi-scale information .

Model ：

It mainly includes a codec network followed by a multi-scale aggregation module , Each layer of coding is convoluted twice ->BN->ReLU, After that maxpool Evacuate to the next floor , The deeper the semantic features extracted, the more helpful it is to correctly identify capillaries ; The low-level features contain rich spatial information, which is helpful to accurately locate the location of capillaries .
 And conventional UNet The difference is that this paper explores a new decoding network to integrate the above information . The decoding part is composed of semantic aggregation modules at different levels , Each aggregation module contains a horizontal connection using spatial information , And contains several skip connection Connect deep semantic information . Because semantic information has better robustness to light and pathological changes , Therefore, with the help of semantic information aggregated at a deep level, the benefits of identifying capillaries and vascular connections can be improved .
 In addition, this paper also introduces a multi-scale aggregation module to solve the problem of large-scale change , By aggregating features of different scales , It is a powerful feature of retinal angiogenesis .

Semantic aggregation module

Including two convolutions and cascading concatination node,concatenation node It will be connected to the coding module of the horizontal same layer , At the same time, it will pass skip connection Connected to all levels of the decoding module . During the connection process, it passes 1x1 Convolution completes the matching of scales ; Cascading features will also pass 1x1 conv Do channel compression , Then two 3x3 Convolution is used to learn more representative features .

Multi scale aggregation module ：
 

  It mainly includes parallel multi branch convolution with different expansion rates to learn information of different scales , The top branch only contains 1x1 Convolution compression channel to d dimension , The latter branch passes 1x1conv Then it will undergo cavity convolution with different expansion rates , Finally, the output of different branches is concatenate And with input Add up add As the output .

Loss function ： In addition to the main loss function Lm Also added 5 Auxiliary loss function L1-L5 In depth Supervision , Therefore, the total loss function is a linear combination of the above loss functions .

experiment ：




 All training data in the experiment go through the following preprocessing steps ：




（1） Gray scale transformation 




（2）CLAHE




（3） Gamma transform 




 The evaluation index ：AUC,PR,ACC.

Experiment 1 ： Ablation Experiment

The baseline model of the results of the ablation experiment is 5 Layer depth UNet Model , Explore the role of semantic aggregation module and multi-scale aggregation module in turn

It can be seen from the ablation implementation that any module is helpful to improve the segmentation performance ; Combine the two modules , This paper can further improve the segmentation performance .

 Experiment two :

Compared with some at present SOTA A comparison of the results , You can see that compared with these methods , The network proposed in this paper has achieved the highest score on these three evaluation indicators of the two data sets .

 

therefore , We can conclude that ： This paper proposes a deep network in retinal vascular segmentation CHASE_DB1 and HRF The most advanced performance has been achieved on public data sets .

summary ：
 The semantic aggregation module effectively aggregates the deep-seated semantic information , Help identify weak capillaries and vascular connections ; The multi-scale aggregation module can effectively combine the multi-scale information , It is helpful to solve the problem of large retinal vascular scale .