(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
Approaches based on the encoder-decoder framework have been widely used in medical image segmentation. However, simply adding or concatenating low-level features and high-level features could be less effective due to the semantic and spatial gaps between them. We propose two approaches for effective feature fusion. One is embedding semantic information into low-level features using a semantic enhancement module (SeEM), the other is introducing spatial details into high-level features using a spatial enhancement module (SpEM). We construct an encoder-decoder network named BGNet towards bridging gaps in multi-level features.
Methods:
The proposed BGNet consists of four major parts: the feature encoder stream, the feature decoder stream, the se-mantic enhancement module (SeEM) and the spatial enhancement module (SpEM). Firstly, input images are fed into the feature encoder stream to extract semantic information. Then the extracted semantic features are fed into the feature decoder stream to recover spatial information. Besides, we use skip connection to fuse low-level features from the encoder stream and high-level features from the decoder stream. We don't concatenate multi-level features directly but first reduce the gaps between them by using SeEM and SpEM. Finally, we use convolution on the last feature to get final segmentation mask.
Results:
We step-wisely decompose our network to investigate the contribution of each proposed component. We use UNet as baseline, then extend the base network by adding SeEM and SpEM respectively. The proposed BGNet achieves significant performance improvement over strong baseline.
Conclusion:
In this paper, we propose two strategies to bridge gaps between multi-level features. One is embedding semantic information into low-level features using a semantic enhancement module (SeEM), the other is introducing spatial details into high-level features using a spatial enhancement module (SpEM). By applying these two modules, we construct an encoder-decoder network named BGNet towards bridging gaps in multi-level features.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the National Natural Science Foundation of China (NNSFC), under Grant Nos. 61375018 and 61672253.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
