Brief Review — MDU-Net: Multi-scale Densely Connected U-Net for Biomedical Image Segmentation

DenseNet Dense Connection Concepts Applied in U-Net

MDU-Net: Multi-scale Densely Connected U-Net for Biomedical Image Segmentation, MDU-Net, by Fudan Unversity, and Univerity of Notre Dame,
2018 arXiv v2, Over 80 Citations (Sik-Ho Tsang @ Medium)
Medical Imaging, Medical Image Analysis, Image Segmentation, U-Net

• MDU-Net, a multiscale densely connected U-Net, is proposed for biomedical image segmentation.
• Three different multi-scale dense connections are proposed for U shaped architectures encoder, decoder and across them.

Outline

1. MDU-Net
2. Results

1. MDU-Net

1.1. Overall Framework

MDU-Net Model Architecture (encoder2-cross3-decoder2)

• The above figure shows MDU-Net (encoder2-cross3-decoder2).
• In contrast to U-Net, there are multiple connections from other levels.
• encoder2 (Purple): 2 means there are 2 connections from previous levels.
• cross3 (Yellow): 3 means there are 3 connections from previous, current, and next levels at the encoder.
• decoder2 (Blue): 2 means there are 2 connections from previous levels, similar to the encoder.

1.2. Encoder

A traditional encoder unit in U-Net VS our purposed dense connected encoder unit

• U-Net (Dashed): Only feature maps from the same level (Dark red) are used for convolution.
• MDU-Net: Feature maps from the previous levels (Light red and red) are downsampled and concatenated with the feature maps from the same level (Dark red) for convolution.

1.3. Decoder

• MDU-Net: Similar to encoder, feature maps from the previous levels (Light red and red) are upsampled and concatenated with the feature maps from the same level (Dark red) for further convolution.

1.4. Cross Connection from Encoder to Decoder

U-Net vs multi-scale dense connections across encoder and decoder unit

• U-Net (Dashed): Only feature maps from the same level (Red) at the encoder are used for concatenation with the feature maps from the same level at the decoder for convolution.
• MDU-Net: Similarly, feature maps from the previous levels (Light red and red) are downsampled, and feature maps from the next levels (Dark red) are upsampled, then they are concatenated with the feature maps from the same level (Red) and feature maps at the decoder (Green) for further convolution.

1.5. Network Quantization

• Incremental Quantization (INQ) is used to compress the parameters as a regularization function against potential overfitting.
• The results of multiple networks are integrated as the final result. The number of parallel model is referring to [39].
• INQ quantizes the parameters to the power of two or zero which makes shift operation possible:

2. Results

2.1. Dataset

• The Gland Segmentation (GlaS) dataset, a biomedical image datasets, in Histology Image Challenge held at MICCAI 2015, is used.
• It contains 165 images with 16 HE stained histological sections colon cancer. 85 images (37 benign and 48 malignant) are selected as training set while 80 images (37 benign and 43 malignant) are used for testing. To be specific, all test images were separated into two categories. (60 Test Part A and 20 Test Part B)
• The network is trained on two NVIDIA GeForce GTX TITAN X, each contains 12 GB of memory. Batch size of 4 is used.

2.2. Ablation Studies

Training loss on the Gland dataset with various dense connected architectures based on U-Net

The proposed models are more stable than original U-Net, which proves that dense connections improve the information flow encoder, decoder and across them.

Prediction performance comparison of U-Net with Multiscale dense connected encoder

Prediction performance comparison of U-Net with Multiscale dense connected decoder

Prediction performance comparison of U-Net with Multiscale dense cross connected block

Each structure is followed by the corresponding number of connections. Concluded from the experiment, obviously, the accuracy generally gets higher as the number of dense connections increases.

Prediction performance comparison of U-Net with Multi-scale dense cross connected block

On MICCAI 2015 Gland Dataset, the modification of certain structure achieves a superiority on average over U-Net.

2.3. Parameter Efficiency

Comparison of parameter number about variant model based U-Net

Due to the reuse of feature maps and concatenation operation, no extra parameters are involved except for the 1×1 convolution.

2.4. SOTA Comparisons

Prediction performance comparison of quantization method

• The overfitting problem is reduced after the first quantization operation in which half of the parameters are quantized. Hence, the performance on Test set B is improved as expected while the prediction accuracy on Test set A remains.

The generalization ability of the model is enhanced compared with the overall quantized model. An surprisingly competitive accuracy of 0.88 is gained on test B. In balance, the half-quantized architecture is adopted as the final model.

2.5. Visualization

Visual gland segmentation results on the GlaS dataset.

Reference

[2018 arXiv v2] [MDU-Net]
MDU-Net: Multi-scale Densely Connected U-Net for Biomedical Image Segmentation

4.2. Biomedical Image Segmentation

2015 … 2018 [MDU-Net] … 2021 [Ciga JMEDIA’21]

My Other Previous Paper Readings