DC-UNet: Rethinking the U-Net Architecture with Dual Channel Efficient CNN for Medical Images Segmen

1 Main Contribution

1. Designed efficient CNN architecture to replace encoder and decoder
2. applied residual module to replace skip connection between encoder and decoder to improve based on state-of-art U-Net model

2 Datasets

• In-house IR breast dataset contains 450 infrared images from 14 patients and 16 healthy volunteers； resolution is 256*128
• EMdataset is 2D EM(电子显微镜图像，公有)：contains 30 images in its training set from a serial section Transmission Electron Microscopy （ssTEM） of the Drosophila(果蝇幼虫EM图像)， resolution is 256*256； Use 5-Fold cross-validation
• Endoscopy images(内窥镜图像 公有): CVC-ClinicDB dataset extracte from the colonoscopy(结肠镜检查)， contains total 612 images; resolution is 128*96

3 Data Augment

1. convert 16-bit to 8-bit and resize 256*128 for thermography breast database
2. resize 256*256 for other databases

4 Initialization

1. All convolutional layers are activated by ReLU
2. Each ReLU later use batch normalization
3. Final output layer activated by Sigmoid
4. Use binary cross-entropy as loss function
5. Use Adam optimizer with the parameter 𝛽1 = 0.9 and 𝛽2 = 0.999

5 Comparison Range

classical U-Net And MultiResUNet
Parameter:

附上一个计算模型参数量的代码：

def count_parameters(model):
    params = [p.numel() for p in model.parameters() if p.requires_grad]
    for item in params:
        print(f'{item:>8}')
    print(f'________\n{sum(params):>8}')

样例结果：

6 Measurement metric

JS(Jaccard similarity), MAE, Tanimoto similarity and SSIM

In conclusion, Jaccard similarity is a proper measure if segmentation output is binary but for grayscale images, Tanimoto similarity is better.

7 Results of Comparison

7.1Results of infrared breast images

• epoch:50
  Results:
  

Because the dataset contains 14 patients and 16 healthy volunteers, the model train dataset in each person respectively

7.2 Results of electron microscopy image

• Use 5-Fold-validation
• epoch: 50
  Results:
  

7.3 Results of endoscopy image

• Use 5-Fold cross-validation
• Epoch: 150
  Results:
  

8 Conclusion

1. it replace the ResBlock in MultiResUNet with DcBlock, which is a two-channel continuous convolutional manipulation.
2. At the same time,it is tested on three datasets with the other models classical UNet and MultiResUNet, and finally DcUNet has the best performance.

9 Core Ways

1. Introduce the classical UNet and MultiResUNet‘s structure
2. List a specific number of channels