Skip to content

Latest commit

 

History

History
219 lines (154 loc) · 13.6 KB

README.md

File metadata and controls

219 lines (154 loc) · 13.6 KB
Raw

Awesome U-Net

Awesome License: MIT

Official repo for Medical Image Segmentation Review: The Success of U-Net

Announcements

August 21, 2024: The final draft is published at the IEEE TPAMI 🔥🔥

November 27, 2022: arXiv release version.

Citation

@article{azad2024medical,
  author={Azad, Reza and Aghdam, Ehsan Khodapanah and Rauland, Amelie and Jia, Yiwei and Avval, Atlas Haddadi and Bozorgpour, Afshin and Karimijafarbigloo, Sanaz and Cohen, Joseph Paul and Adeli, Ehsan and Merhof, Dorit},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Medical Image Segmentation Review: The Success of U-Net}, 
  year={2024},
  pages={1-20},
  keywords={Image segmentation;Biomedical imaging;Taxonomy;Computer architecture;Feature extraction;Transformers;Task analysis;Medical Image Segmentation;Deep Learning;U-Net;Convolutional Neural Network;Transformer},
  doi={10.1109/TPAMI.2024.3435571}
}

Abstract

Automatic medical image segmentation is a crucial topic in the medical domain and successively a critical counterpart in the computer-aided diagnosis paradigm. U-Net is the most widespread image segmentation architecture due to its flexibility, optimized modular design, and success in all medical image modalities. Over the years, the U-Net model achieved tremendous attention from academic and industrial researchers. Several extensions of this network have been proposed to address the scale and complexity created by medical tasks. Addressing the deficiency of the naive U-Net model is the foremost step for vendors to utilize the proper U-Net variant model for their business. Having a compendium of different variants in one place makes it easier for builders to identify the relevant research. Also, for ML researchers it will help them understand the challenges of the biological tasks that challenge the model. To address this, we discuss the practical aspects of the U-Net model and suggest a taxonomy to categorize each network variant. Moreover, to measure the performance of these strategies in a clinical application, we propose fair evaluations of some unique and famous designs on well-known datasets. We provide a comprehensive implementation library with trained models for future research. In addition, for ease of future studies, we created an online list of U-Net papers with their possible official implementation.

The structure of codes

Here is

.
├── README.md
├── images
│   └── *.png
├── configs
│   ├── isic
│   │   ├── isic2018_*<net_name>.yaml
│   └── segpc
│       └── segpc2021_*<net_name>.yaml
├── datasets
│   ├── *<dataset_name>.py
│   ├── *<dataset_name>.ipynb
│   └── prepare_*<dataset_name>.ipynb
├── models
│   ├── *<net_name>.py
│   └── _*<net_name>
│       └── *.py
├── train_and_test
│   ├── isic
│   │   ├── *<net_name>-isic.ipynb
│   │   └── *<net_name>-isic.py
│   └── segpc
│       ├── *<net_name>-segpc.ipynb
│       └── *<net_name>-segpc.py
├── losses.py
└── utils.py

Dataset prepration

Please go to "./datasets/README.md" for details. We used 3 datasets for this work. After preparing required data you need to put the required data path in relevant config files.

Train and Test

In the train_and_test folder, there are folders with the names of different datasets. In each of these subfolders, there are files related to each model network in two different formats (.py and ‍.ipynb). In notebook files you will face with the following procedures. This file contains both the testing and traning steps.

  • Prepration step
    • Import packages & functions
    • Set the seed
    • Load the config file
  • Dataset and Dataloader
    • Prepare Metrics
  • Define test and validate function
  • Load and prepare model
  • Traning
    • Save the best model
  • Test the best inferred model
    • Load the best model
  • Evaluation
    • Plot graphs and print results
  • Save images

Pretrained model weights

Here you can download pre-trained weights for networks.

Network Model Weight Train and Test File
U-Net ISIC2018 - SegPC2021 ISIC2018 - SegPC2021
Att-UNet ISIC2018 - SegPC2021 ISIC2018 - SegPC2021
U-Net++ ISIC2018 - SegPC2021 ISIC2018 - SegPC2021
MultiResUNet ISIC2018 - SegPC2021 ISIC2018 - SegPC2021
Residual U-Net ISIC2018 - SegPC2021 ISIC2018 - SegPC2021
TransUNet ISIC2018 - SegPC2021 ISIC2018 - SegPC2021
UCTransNet ISIC2018 - SegPC2021 ISIC2018 - SegPC2021
MISSFormer ISIC2018 - SegPC2021 ISIC2018 - SegPC2021

Results

For evaluating the performance of some mentioned methods, three challenging tasks in medical image segmentaion has been considered. In bellow, results of them illustrated.


Performance comparison on ISIC 2018 dataset (best results are bolded).

Methods AC PR SE SP Dice IoU
U-Net 0.9446 0.8746 0.8603 0.9671 0.8674 0.8491
Att-UNet 0.9516 0.9075 0.8579 0.9766 0.8820 0.8649
U-Net++ 0.9517 0.9067 0.8590 0.9764 0.8822 0.8651
MultiResUNet 0.9473 0.8765 0.8689 0.9704 0.8694 0.8537
Residual U-Net 0.9468 0.8753 0.8659 0.9688 0.8689 0.8509
TransUNet 0.9452 0.8823 0.8578 0.9653 0.8499 0.8365
UCTransNet 0.9546 0.9100 0.8704 0.9770 0.8898 0.8729
MISSFormer 0.9453 0.8964 0.8371 0.9742 0.8657 0.8484

Performance comparison on SegPC 2021 dataset (best results are bolded).

Methods AC PR SE SP Dice IoU
U-Net 0.9795 0.9084 0.8548 0.9916 0.8808 0.8824
Att-UNet 0.9854 0.9360 0.8964 0.9940 0.9158 0.9144
U-Net++ 0.9845 0.9328 0.8887 0.9938 0.9102 0.9092
MultiResUNet 0.9753 0.8391 0.8925 0.9834 0.8649 0.8676
Residual U-Net 0.9743 0.8920 0.8080 0.9905 0.8479 0.8541
TransUNet 0.9702 0.8678 0.7831 0.9884 0.8233 0.8338
UCTransNet 0.9857 0.9365 0.8991 0.9941 0.9174 0.9159
MISSFormer 0.9663 0.8152 0.8014 0.9823 0.8082 0.8209

Performance comparison on Synapse dataset (best results are bolded).

Method DSC↑ HD↓ Aorta Gallbladder Kidney(L) Kidney(R) Liver Pancreas Spleen Stomach
U-Net 76.85 39.70 89.07 69.72 77.77 68.60 93.43 53.98 86.67 75.58
Att-UNet 77.77 36.02 89.55 68.88 77.98 71.11 93.57 58.04 87.30 75.75
U-Net++ 76.91 36.93 88.19 65.89 81.76 74.27 93.01 58.20 83.44 70.52
MultiResUNet 77.42 36.84 87.73 65.67 82.08 70.43 93.49 60.09 85.23 74.66
Residual U-Net 76.95 38.44 87.06 66.05 83.43 76.83 93.99 51.86 85.25 70.13
TransUNet 77.48 31.69 87.23 63.13 81.87 77.02 94.08 55.86 85.08 75.62
UCTransNet 78.23 26.75 84.25 64.65 82.35 77.65 94.36 58.18 84.74 79.66
MISSFormer 81.96 18.20 86.99 68.65 85.21 82.00 94.41 65.67 91.92 80.81

Visualization

  • Results on ISIC 2018

    isic2018.png

    Visual comparisons of different methods on the ISIC 2018 skin lesion segmentation dataset. Ground truth boundaries are shown in green, and predicted boundaries are shown in blue.

  • Result on SegPC 2021

    segpc.png

    Visual comparisons of different methods on the SegPC 2021 cell segmentation dataset. Red region indicates the Cytoplasm and blue denotes the Nucleus area of cell.

  • Result on Synapse

    synapse.png

    Visual comparisons of different methods on the Synapse multi-organ segmentation dataset.

References

Codes [GitHub Pages]

Query

For any query, please contact us.

rezazad68@gmail.com
engtekh@gmail.com
afshinbigboy@gmail.com