Contains the prediction codes for our submission to the Foot Ulcer Segmentation Challenge at MICCAI 2021 that placed us in the 1st rank in the challenge legacy leaderboard.
If you find the contents of this repository useful or use the provided codes, please use the following citation:
BibTex entry:
@INPROCEEDINGS{9956253,
author={Mahbod, Amirreza and Schaefer, Gerald and Ecker, Rupert and Ellinger, Isabella},
booktitle={2022 26th International Conference on Pattern Recognition (ICPR)},
title={Automatic Foot Ulcer Segmentation Using an Ensemble of Convolutional Neural Networks},
year={2022},
volume={},
number={},
pages={4358-4364},
doi={10.1109/ICPR56361.2022.9956253}}
Paper link: https://ieeexplore.ieee.org/abstract/document/9956253
Preprint link: https://arxiv.org/abs/2109.01408
.
├── code
│ ├── test.py (main run file)
│ ├── gpu_setting.py
│ ├── metric.py
│ └── params_test.py (all configs are here)
│
├── saved_models
│ ├── linknet(LinkNet models should be downloaded from Google drive and placed in this folder)
│ └── unet (U-Net models should be downloaded from Google drive and placed in this folder)
│
├── test images (place all test images here (size 512x512 pixels))
│
└── results
├── temp
└── final (final results will be saved here)
1- download the saved models from Google Drive: https://drive.google.com/drive/folders/1lprFVD--hzFLglXpe8di0CkqSXRK2DFO?usp=sharing and place them in saved_models folder
2- put the test images inside the test_images folder (already included 200 test images from: https://github.com/uwm-bigdata/wound-segmentation/tree/master/data/Foot%20Ulcer%20Segmentation%20Challenge/test/images but you could add/remove images)
3- build the Docker environment
docker build -f Dockerfile -t fuseg2021_amirreza_mahbod_medicaluniversityofvienna .
or download the built image from:https://drive.google.com/file/d/1K4j9gXKzmLfLAe-LjAhRLIi1Vz-_ghkp/view?usp=sharing
and extract the .tar file.
docker load --input fuseg2021_amirreza_mahbod_medicaluniversityofvienna.tar
4- run the container
docker run --gpus all -v /home/masih/Desktop/wound_docker/results/:/src/results/ -ti fuseg2021_amirreza_mahbod_medicaluniversityofvienna /bin/bash
note: you need to chenge /home/masih/Desktop/wound_docker/results/ to the path that you want to save the results on your local system
5- run the following commands inside the container:
$ cd src/code
$ python3 test.py
6- final results will be saved inside results/final folder
To derive the results in the following table, we used the Medetec foot ulcer dataset [1] for pre-training. Then we used the training set of the MICCAI 2021 Foot Ulcer Segmentation Challenge dataset [2] (810 images) as the training set. The reported results in the following table are based on the validation set of the Foot Ulcer Segmentation dataset (200 images). For the challenge submssion, we used the entire 1010 images of the train and validation set to train our models.
| Model | Image-based Dice (%) | Precision (%) | Recall (%) | Dataset-based IOU (%) | Dataset-based Dice (%) |
|---|---|---|---|---|---|
| VGG16 [2] | - | 83.91 | 78.35 | - | 81.03 |
| SegNet [2] | - | 83.66 | 86.49 | - | 85.05 |
| U-Net [2] | - | 89.04 | 91.29 | - | 90.15 |
| Mask-RCNN [2] | - | 94.30 | 86.40 | - | 90.20 |
| MobileNetV2 [2] | - | 90.86 | 89.76 | - | 90.30 |
| MobileNetV2 + pp [2] | - | 91.01 | 89.97 | - | 90.47 |
| EfficientNet1 LinkNet (this work) | 83.93 | 92.88 | 91.33 | 85.35 | 92.09 |
| EfficientNet2 U-Net (this work) | 84.09 | 92.23 | 91.57 | 85.01 | 91.90 |
| Ensemble U-Net LinkNet (this work) | 84.42 | 92.68 | 91.80 | 85.51 | 92.07 |
Amirreza Mahbod (Medical University of Vienna)
Email: amirreza.mahbod@meduniwien.ac.at
[1] Thomas, S. Stock pictures of wounds. Medetec Wound Database (2020). http://www.medetec.co.uk/files/medetec-image-databases.html
[2] Wang, C., Anisuzzaman, D.M., Williamson, V. et al. Fully automatic wound segmentation with deep convolutional neural networks. Sci Rep 10, 21897 (2020). https://doi.org/10.1038/s41598-020-78799-w