Brief Review — RCU-Net: Tumor Segmentation in Breast Ultrasound Image

RCU-Net, Improve Skip Connections Using Residual Blocks and Dense Block

Sik-Ho Tsang
3 min readMar 12, 2023
Original ultrasound images. (a) examples of fibroadenoma; (b) examples of invasive ductal carcinoma; (c) examples of lobular hyperplasia.

Tumor Segmentation in Breast Ultrasound Image by Means of Res Path Combined with Dense Connection Neural Network,
RCU-Net, by University of Shanghai for Science and Technology,
2021 MDPI J. Diagnostics, Over 5 Citations (

@ Medium)
Medical Imaging, Medical Image Analysis, Image Segmentation, U-Net

4.2. Biomedical Image Segmentation
2015–2021 [Expanded U-Net] [3-D RU-Net] [nnU-Net] [TransUNet] [CoTr] [TransBTS] [Swin-Unet] [Swin UNETR] 2022 [UNETR]
My Other Previous Paper Readings Also Over Here

  • An improved U-Net, namely RCU-Net, is proposed in this paper:
  1. First, Res Path was introduced into the U-Net to reduce the difference between the feature maps of the encoder and decoder.
  2. Then, a new connection, dense block, originated from DenseNet, is introduced at the top level of U-Net as well, was added to reduce the feature information loss and alleviate the vanishing gradient problem.

Outline

  1. RCU-Net
  2. Results

1. RCU-Net

1.1. Preprocessing

Preprocessing
  • The Contrast Limited Adaptive Histogram Equalization (CLAHE) is used to accentuates the contrast between the tumor and surrounding tissue.

After CLAHE, the range of gray values is much larger, which is helpful for deep learning model training.

1.2. Overall Architecture

Overall Architecture

Res Path is proposed at the skip connections of U-Net.

Also, Dense block, originated from DenseNet, is proposed to add the input to the near end of network.

1.3. Res Path

Res Path

Two parallel 3×3 and 1×1 convolutional layers with BN and ReLU are used for each Res block, which is used to reduce the semantic gap between encoder and decoder.

  • And there is 2 Res Blocks at the above example.

1.4. Dense Block

Dense Block

A new connection is introduced that incorporates a dense block from the input of the feature maps in the encoding-to-decoding component.

2.1. Ablation Study

The mean and standard deviation of the five indexes using 10-fold cross validation. The numbers behind the mean are the standard deviation
  • 10-fold cross validation is used for evaluation.

Compared with that of the U-Net, the TP, JS, and DC of the proposed model improved by 1.0625%, 2.135%, and 2.0105%, respectively. The HD index decreased by 1.41023.

Visualizations for Different Components Enabled

U-Net had lesser number of pixels false predicted due to Res Path.

2.2. SOTA Comparisons

Comparison of different segmentation models.

The results show that the U-Net, DIU-Net, and the proposed model show better performance compared with SegNet and FCN-8s because of the skip connection.

Additionally, the DIU-Net had close performance to the proposed model due to the deeper layer. However, excessive layers may lead to too many parameters, making the model more difficult and slower to train.

An unsuccessful segmentation case. (a) Original figure; (b) Ground Truth obtained from (a); (c) segmentation result of the U-Net model; (d) segmentation result of the U-Net with Res Path model; (e) segmentation result of the proposed model.

Though the model performed better than others, there were still some failed segmentation cases owing to the difficulty of extracting features in some images.

--

--

Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: https://linktr.ee/shtsang for Twitter, LinkedIn, etc.