Brief Review — DIU-Net: DENSE-INception U-Net for Medical Image Segmentation

DIU-Net, Using DenseNet and Inception-v3 Concepts in U-Net

DENSE-INception U-Net for Medical Image Segmentation,
DIU-Net, by Northeastern University, and Ulster University,
2020 J. CMPB, Over 120 Citations (Sik-Ho Tsang @ Medium)
Medical Imaging, Medical Image Analysis, Image Segmentation

• DENSE-INception U-Net (DIU-Net) is proposed, that integrates the Inception-Res module and densely connecting convolutional module into the U-Net architecture.

Outline

1. DENSE-INception U-net (DIU-Net)
2. Results

1. DENSE-INception U-net (DIU-Net)

1.1. Overall Architecture

Overall architecture of the DIU-Net model.

• DENSE-INception U-Net (DIU-Net) is proposed as shown above.
• This network model contains the analysis path and the synthesis path, and these two paths mainly consist of four kinds of blocks, namely the Inception-Res block, the Dense-Inception block, the down-sample block and the up-sample block.
• The analysis path consists of three Inception-Res blocks, one Dense-Inception block and four down-sample blocks.
• The synthesis path consists of three Inception-Res blocks, one Dense-Inception block and four up-sample blocks.
• In the middle of the network, a single Dense-Inception block is deployed and this block contains much more inception layers than the others.

1.2. Inception-Res Block

Overall architecture of the Inception-Res block.

• The main purpose is to aggregate feature maps from different branches of kernels of different sizes, which can make the network wider and capable of learning more features, as above.
• Different from the original Inception-Res architecture as in Inception-v3, each convolutional layer is followed by a batch normalization (BN) layer except for bottleneck layers.

1.3. Dense-Inception block

Left: Overall architecture of a 4-layer Dense-Inception block. Right: Overall architecture of each Inception-Res block in Dense-Inception block.

• As seen above, DenseNet concept is applied.
• Three Dense-Inception blocks (Left) are designed in total: one block is set in the analysis path, one is in the synthesis path, and the last one is set in the middle of network.
• Each Dense-Inception block except the middle one contains 12 proposed Inception-Res modules (Right), and the middle one has 24 Inception-Res modules.

1.3. Down-sample & Up-sample Blocks

The down-sample block (Left) and the up-sample block (Right).

• They are simplified inception module with three branches.

2. Results

2.1. Lung Segmentation

Experimental outputs for lung dataset using different kinds of methods.

Experimental results of proposed approaches for lung segmentation and comparison against other networks.

DIU-Net network shows better performance under each evaluation index.

2.2. Retina Blood Vessel Segmentation

Experimental outputs for retina blood vessel segmentation using different kinds of methods.

Experimental results of proposed approaches for retina blood vessel and comparison against other networks or traditional methods.

DIU-Net model shows better performance in terms of the DICE coefficient, Jaccard similarity, accuracy, F1-score and the sensitivity. Compared with the traditional methods, the DIU-Net also achieves better performance.

2.3. Brain Tumor Segmentation

Experimental outputs for brain tumor segmentation using different kinds of methods.

Experimental results of proposed approaches for brain tumor segmentation and comparison against other networks or traditional methods.

DIU-Net network shows better performance under each evaluation index.

2.4. Runtime

Training/Testing time on various kinds of models.

DIU-Net is slower than other networks, considering the improved performance, its running time is acceptable.

Reference

[2020 J. CMPB] [DIU-Net]
DENSE-INception U-Net for Medical Image Segmentation

2015–2020 [MultiResUNet] [UNet 3+] [Dense-Gated U-Net (DGNet)] [Non-local U-Net] [SAUNet] [DIU-Net] 2021 [Expanded U-Net]

==== My Other Previous Paper Readings ====