giser18 / 1d-2d-segmentation-autoencoder-tf2-keras

1D and 2D Segmentation Models with options such as Deep Supervision, Guided Attention, BiConvLSTM, Autoencoder, etc.

Segmentation-Model-Builder-Tensorflow-Keras

This repository contains 1D and 2D Signal Segmentation Model Builder for UNet, several of its variants and other models developed in Tensorflow-Keras. The code supports Deep Supervision, Autoencoder mode, Guided Attention, Bi-Directional Convolutional LSTM and other options explained in the codes and demos. The segmentation models can be used for binary or multiclass segmentation, or for regression tasks.

Models supported

1. UNet [1]
2. UNet Ensembled (UNetE) [2]
3. UNet+ (UNetP) [2]
4. UNet++ (UNetPP) [3]
5. UNet3+ []
6. MultiResUNet [4]
7. DenseInseptionUNet []
8. Feature Pyramid Network (FPN) []
9. Pyramid Spatial Pooling (PSP) Net []
10. AlbUNet []
11. TernausUNet []
12. BCDUNet []
13. SEDUNet or MGCUNet []
14. IBAUNet []

UNet to UNet++

As it can be seen, from UNet to UNet++, the models become more nested and denser (so heavier i.e., more parameters). UNet++ is like a combination of UNetE and UNet+. Normally users do not try UNet+ or UNetE. But for some problems, UNet++ might overfit. In those cases, these intermediate models can be tried instead of using a shallower or narrower UNet++ model. UNet to UNet++ Transformation Process

MultiResUNet

MultiResUNet has an interesting Residual path for the skip connection and uses MultiRes Blocks instead of normal CNN blocks for deep learning. MultiResUNet model also uses Transposed Convolutions in the encoder by default instead of UpSampling.
MultiResUNet Architecture

Internal structure for the MultiResUNet block is shown below [4]. As it can be seen, the MultiRes Block goes through convolutions of different Kernel sizes, then concatenates in the end.
MultiRes Block of MultiResUNet

Structure for the Residual Path (or ResPath) is shown below [4]. The Residual Path has multiple covolutions and additions over the skip conncetion, in place of UNet's direct skip connection or Residual path which gets concatenated directly.
Residual Path of MultiResUNet

Supported Features

The speciality about this model is its flexibility, such as:

1. The user can choose any of the 5 available UNet variants for either 1D or 2D Segmentation tasks.
2. The models can be used for Binary or Multi-Class Classification, or Regression type Segmentation tasks.
3. The models allow Deep Supervision [5] with flexibility during Segmentation.
4. The segmentation models can also be used as Autoencoders [6] for Feature Extraction.
5. The segmentation models can also be implemented with Guided Attention [8].
6. The depth of all the models can be varied to form very shallow to very deep networks.
7. Number of input kernel/filter, commonly known as Width of the model can be varied.
8. Number of classes for Classification tasks and number of extracted features for Regression tasks can be varied.
9. Number of Channels in the Input Dataset can be varied.
10. In case of only MultiResUNet, its 'alpha' parameter can be varied (default set to 1.0) [4].
11. All the networks use Transposed Convolution [7] by default (instead of UpSampling) in the Decoder section, which can be turned off to use traditional UpSampling.

Transposed Convolution [7]

Attention Guided UNet [8]

Details of the process are available in the DEMO provided in the codes section. The datasets used in the DEMO as also available in the 'Documents' folder. [The DEMO will be added soon for 1D and 2D]

References

[1] Ronneberger, O., Fischer, P., & Brox, T. (2021). U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv.org. Retrieved 30 August 2021, from https://arxiv.org/abs/1505.04597.
[2] Zhou, Z., Siddiquee, M., Tajbakhsh, N., & Liang, J. (2021). UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation. Arxiv-vanity.com. Retrieved 30 August 2021, from https://www.arxiv-vanity.com/papers/1912.05074/.
[3] Zhou, Z., Siddiquee, M., Tajbakhsh, N., & Liang, J. (2021). UNet++: A Nested U-Net Architecture for Medical Image Segmentation. arXiv.org. Retrieved 30 August 2021, from https://arxiv.org/abs/1807.10165.
[4] Ibtehaz, Nabil, and M. Sohel Rahman. “Multiresunet : Rethinking the u-Net Architecture for Multimodal Biomedical Image Segmentation.” ArXiv.org, 11 Feb. 2019, arxiv.org/abs/1902.04049v1.
[5] Wang, L., Lee, C., Tu, Z., & Lazebnik, S. (2021). Training Deeper Convolutional Networks with Deep Supervision. arXiv.org. Retrieved 30 August 2021, from https://arxiv.org/abs/1505.02496.
[6] Chang, H. (2021). A Method of Brain Image Optimization based on an Autoencoder Unet. Journal Of Physics: Conference Series, 1952(2), 022064. https://doi.org/10.1088/1742-6596/1952/2/022064
[7] Transposed Convolution Demystified. Medium. (2021). Retrieved 1 September 2021, from https://towardsdatascience.com/transposed-convolution-demystified-84ca81b4baba.
[8] Li, K., Wu, Z., Peng, K., Ernst, J. and Fu, Y., 2021. Tell Me Where to Look: Guided Attention Inference Network. [online] arXiv.org. Available at: https://arxiv.org/abs/1802.10171.