In this story, MultiResUNet, by Samsung, and Bangladesh University of Engineering and Technology, is briefly reviewed. By enhancing the U-Net architecture, MultiResUNet outperforms U-Net on five datasets. This is a paper in 2020 Journal of Neural Networks with high impact factor of 5.785. (Sik-Ho Tsang @ Medium)
Outline
- MultiResUNet: Modifications on U-Net
- Experimental Results
1. MultiResUNet: Modifications on U-Net
1.1. From 2 Conv Layers in U-Net to MultiRes Block in MultiResUNet
- For the sequence of two convolutional layers at each level in the original U-Net, they are replaced by the proposed MultiRes block.
- (a): First, start with a simple Inception-like block by using 3×3, 5×5 and 7×7 convolutional filters in parallel, to reconcile spatial features from different context size.
- (b): Then, large filter is factorized into a succession of 3 × 3 filters.
- (c): Finally, MultiRes block is established, by increasing the number of filters in the successive three layers gradually and adding a residual connection, along with 1×1 filters for conserving dimensions.
- This is similar to the DenseBlock in DenseNet with the residual path, originated in ResNet.
1.2. ResPath in MultiResUNet
- For the ResPath, there are 3×3 and 1×1 filters as shown above. Number of 3×3 and 1×1 filters depends on the level inside the network, which as shown at the table below.
- Authors hypothesized that the intensity of the semantic gap between the encoder and decoder feature maps are likely to decrease.
- These additional non-linear operations are expected to reduce the semantic gap between encoder and decoder features.
- (Yet, according to the figure above, is it still a residual path? lol)
1.3. Number of feature maps
- In order for fair comparison with U-Net, similar number of parameters should be maintained between two models:
- where U and W are the number of filters in one convolutional layer in U-Net and MultiResUNet respectively. α = 1.67 is used. Thus, the filter numbers as shown below are multipled by α already.
1.4. Architecture Summary
- Below is the details of MultiResUNet architecture:
2. Experimental Results
2.1. Datasets
- 5 Datasets are used as shown above. The first 4 are 2D images and the last one are 3D images.
2.2. Model Sizes
- As shown above, the parameters of U-Net and MultiResUNet models for 2D and 3D data are similar for fair comparison.
2.3. Evaluation Metric
- Jaccard Index is used as evaluation metric, similar to IoU.
2.4. Results
- MultiResUNet obtains much higher Jaccard Index than U-Net.
2.5. Qualitative Results
- With such subtle boundary, MultiResUNet can segment with much higher accuracy.
During the days of coronavirus, let me have a challenge of writing 30 stories again for this month ..? Is it good? This is the 19th story in this month. Thanks for visiting my story..
Reference
[2020 JNN] [MultiResUNet]
MultiResUNet : Rethinking the U-Net Architecture for Multimodal Biomedical Image Segmentation
Biomedical Image Segmentation
[CUMedVision1] [CUMedVision2 / DCAN] [U-Net] [CFS-FCN] [U-Net+ResNet] [MultiChannel] [V-Net] [3D U-Net] [M²FCN] [SA] [QSA+QNT] [3D U-Net+ResNet] [Cascaded 3D U-Net] [Attention U-Net] [RU-Net & R2U-Net] [VoxResNet] [DenseVoxNet][UNet++] [H-DenseUNet] [DUNet] [MultiResUNet]