MDU-Net: Multi-scale Densely Connected U-Net for ...

1 MDU-Net: Multi-scale Densely Connected U-Net forbiomedical image segmentationJiawei ZhangFudan JinFudan XuFudan XuUniverity of Notre ZhangFudan is doctor s doctor , biomedical image seg-mentation plays a central role in quantitative analysis, clin-ical diagnosis, and medical intervention. In the light of thefully convolutional networks (FCN) and U-Net , deep con-volutional networks (DNNs) have made significant contri-butions in biomedical image segmentation applications. Inthis paper, based on U-Net , we propose MDUnet, a Multi-scale Densely Connected U-Net for biomedical image seg-mentation. we propose three different Multi-scale denseconnections for U shaped architectures encoder, decoderand across them.

2 The highlights of our architecture is di-rectly fuses the neighboring different scale feature mapsfrom both higher layers and lower layers to strengthen fea-ture propagation in current layer. Which can largely im-proves the information flow encoder, decoder and acrossthem. Multi-scale dense connections, which means con-taining shorter connections between layers close to the in-put and output, also makes much deeper U-Net adopt the optimal model based on the experiment andpropose a novel Multi-scale Dense U-Net (MDU-Net) ar-chitecture with quantization. Which reduce overfitting inMDU-Net for better accuracy.

3 We evaluate our purposemodel on the MICCAI 2015 Gland Segmentation dataset(GlaS). The three Multi-scale dense connections improve U-Net performance by up to on test A and on testB in the MICCAI Gland dataset. Meanwhile the MDU-netwith quantization achieves the superiority over U-Net per-formance by up to 3% on test A and on test IntroductionBiological structures to support medical diagnosis, sur-gical planning and treatments. Based on fully convolu-Figure 1. Example of a Multi-scale dense Connected encoder networks (FCN) and U-Net [31, 26], deep convolu-tional networks (DNNs) have made significant improvem-nents in biomedical image segmentation.

4 Due to the highefficiency and capability to automatically capture informa-tion without hand-designed features, deep learning meth-ods have dominated biomedical image analysis. Due tothe segmentation abnormalities and histological variations,a higher level of pixelwise prediction in biomedical imageanalysis is required than in natural images. In particular,a marginal bias in biomedical segmentation will result inhigh false clinical treatment. Therefore, the improvementof segmentation remains boosting attention. Recent workssuch as U-Net which applied skip connections to combinefeature maps from the current layer with higher layer fea-ture maps and proved a competitive performance in main-taining fine-grained information.

5 In the meantime, segmen-tation masks are generated with contextual details even ifthe background composition is rather complicated. We di-vide it into two )intra-block dense connectionswhich embeds the dense block to the traditional convolu-tional block such as FDU-Net[15]. In addition, cascaded ofstacked U-Nets also gain enough attention. CU-Net[9] per-1 [ ] 4 Dec 2018form dense connections of the same level among multipleU-Nets. However, these works fail to consider transform-ing the size of feature maps. As a consequence, they aresubstantially different from our )Inter-block denseconnections. Which means current layer can fuses fromprevious layer with differnet scale .

6 For instance, MIMO-Net[30] takes input image of different scales in the encoderunit. However, the feature maps are not actually reused. U-Net ++[45] fuses higher resolution feature maps in the de-coder unit but it involves a massively computational costsdue to the large number of intermediate convolutions. InU-Net ++, the current layer can only fuse the feature mapsfrom higher by DenseNet[18], in order to improve segmen-tation accuracy, we directly down-sample features fromlower layers and perform up-sampling functions for higherlayers to the same resolution of the current layer and fusesthem with feature maps from the current layer.

7 We use 1*1conv twice to control the number of channels the same asbefore. The whole operation involves in a small constantnumber of extra parameters. As far as we are concerned,we are the first to explore directly fusing deep semantic andcoarse-grained feature maps from higher layers and low-level, fine-grained feature maps from lower layers. Themodified fusing operation contains more object informationand pixel information, and therefore improves the segmen-tation in U-Net architecture. We also systematically analyzethe impact of different kinds of Densely Connected experiment shows that fusing higher and lower layer sfeature maps simultaneously turns out more effective andachieves a higher contribution of our work is1)conducting completeexperiment and analysis on the influence on U-Net withmulti- scale dense connections )we adoptthe optimal model based on the experiment and proposea novel Multi-scale Dense U-Net (MDU-Net) architecturewith quantization.

8 The proposed model achieves the supe-riority over U-Net performance by up to 3% on testA on Related WorkIn this section, we introduce late approaches towards U-Net architecture, dense connections, Multi-scale representa-tion, network quantization and biomedical image segmenta-tion U-Net architectureModels are designed as encoder-decoder architectures toretrieve high resolution from low resolution representationsof the image. [31] initially proposed the U-shape networkarchitecture with direct skip connection between the en-coder and decoder. systematically analyzed and proved theimportance of long skip connection in U-Net for biomed-ical image than image segmenta-tion, a variety of tasks involves in U-Net based architec-ture.

9 Stacked U-Nets[34] iteratively fuse Multi-scale fea-tures without changing the resolutions. To deal with hu-man pose estimation tasks, [27, 38, 41] stacked modifiedU-Nets which captured both the top-down and bottom-upfeatures as a whole. [33][12]follow the grid pattern in theU-shape structure. In a more general manner, [24] addition-ally employed multi -path refinement and global convolu-tional blocks respectively between the encoder and classification and localization problems are solved si-multaneously during the successive down-sampling and up-sampling operation in U-Net . Furthermore, we conduct ex-periments in detail on the impact of U-Net architecture withvarious dense Dense connectionsRecently, the exploration on both the depth and thewidth of the network architecture has been a focusedstudy.

10 Approaches toward wider network begin with [36,37]which introduced Inception Module by concatenatingfeature maps to approximate sparse structure. Moreover,residual network [17, 19]alleviated the vanishing gradientproblem by summing up a shortcut connection with theresidual function. Recent methods such as PSPNet [43] andRefinenet [24] applied residual architecture more frequentlyas feature extractor in dense prediction tasks. [11]combinedU-Net with residual network and proved skip connectioneffective in qiomedical image segmentation. Additionally,to improve the representational power without increasingthe depth and width of the network, [18] proposed a typicalstructure of dense connections.