Brief Review — PRDNet: Medical Image Segmentation Based on Parallel Residual and Dilated Network

PRDNet, Parellel Residual & Dilation Paths

Sik-Ho Tsang
3 min readMar 16


PRDNet: Medical Image Segmentation Based on Parallel Residual and Dilated Network,
PRDNet, by Hebei University of Technology,
2021 Elsevier J. Measurement, Over 10 Citations (Sik-Ho Tsang @ Medium)
Medical Imaging, Medical Image Analysis, Image Segmentation

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

  • PRDNet (Parallel Residual and Dilated Network) is proposed, where ResNet and dilated convolution are simultaneously used to extract multilayer features of medical images in parallel.
  • In the decoding stage, the multi-layer features are fused according to the structure of feature pyramid network (FPN).


  1. PRDNet
  2. Results

1. PRDNet

PRDNet Model Architecture

1.1. Backbone

  • The input image size of PRDNet is 256×256.
  • ResNet can be divided into five stages according to the size and the channels of the feature map.
  • The first stage of ResNet consists of convolution and pooling, which down-samples the input image to 64 × 64. The stages from the second to the fifth are called 𝐶2, 𝐶3, 𝐶4 and 𝐶5, respectively.
  • In ResNet-101, 𝐶2 does not change the size of the image but increases the channels of the feature map. Besides, the images are down-sampled at each stage starting with 𝐶3. Subsequently, the size of the image changes to 8×8.

In PRDNet, ResNet and dilated convolution share the 𝐶2 and 𝐶3 layers. After that, they are in a parallel way.

  • Different with ResNet, the last two layers of dilated convolution are called 𝐶4𝑑 and 𝐶5𝑑 and they are both 32 × 32 in size.

At Last, ResNet and dilated convolution produce a total of six-layer features for subsequent fusion.

1.2. Fusion

  • Feature pyramid is a bottom-up and top-down structure composed of multi-level features.
  • At dilation branch, 𝐶4𝑑 and 𝐶5𝑑 are added to the up-sampled results from 𝐶5 layer and convolved respectively to obtain 𝑃4𝑑 and 𝑃5𝑑.
  • At ResNet branch, 𝑃2-𝑃5 layers are obtained by up-sampling layer by layer according to the original FPN structure. Each layer is convolved, up-sampled, and added together.
  • In order to get the final segmentation result, the output feature map is up-sampled to the same size as the input image and convolved to make the output channels equal to the number of labels.

1.3. Loss Function

  • Cross-entropy loss is used:

2. Results

Results of Different Algorithms on CHAOS Dataset (Left) and ISIC2017 Dataset (Right).

PRDNet achieves the best performance on the two datasets.

Results of different algorithms on images from four patients. Each image contains one or more small objects to be segmented.
Heatmaps of different organs generated by different algorithms. (a), (b), (c) and (d) represent liver, right kidney, left kidney and spleen, respectively.

It is clear that PRDNet, Attention U-Net (A-UNet), DANet and SENet perform better at detecting target locations. The segmentation result of Attention U-Net (A-UNet) is inferior to PRDNet in detail.



Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: for Twitter, LinkedIn, etc.