In this paper, an automated deep learning-based framework for bone shadow suppression from frontal CXRs is developed. The framework was inspired by U-Net-based convolutional neural networks (CNNs). Multiple deep learning-based CNN architectures were implemented to fulfill this task. Among those, a novel neural network architecture called xU-NetFullSharp was proposed. This network is inspired by the most modern U-NetSharp [8] architecture and combines different approaches to preserve as many details, as possible and accurately suppress bone shadows. Additionally, recent state-of-the-art CNN models from [9] and [3] designed for this task were used for comparison. Code for the utilized models is available in the models folder in this repository.
This guide provides an overview of setting up and testing bone suppression models using Conda and Python scripts.
To create the Conda environment with necessary dependencies, run:
conda env create -f environment.ymlAfter creating the environment, activate it by executing:
conda activate bone_suppressionYou can test individual models on either internal or external datasets. To do this, use the test.py script with the following command:
python test.py --model_name <desired_model> --test_variant <external | internal> --data_path <path_to_desired_dataset>Supported Models:
U-Net:"UNET"[10],"ATT_UNET"[7],"DEEP_RESUNET"[11]U-Net++:"UNETPP"[12],"ATT_UNETPP"[5]U-Net3+:"UNET3P"[2]U-Net#:"UNET_SHARP"[8]xU-NetFullSharp(Ours):"XUNETFS","ATT_XUNETFS"Kalisz-Marczyk Autoencoder[3]:"KALISZ_AE"DeBoNet[9]:"UNET_RES18","FPN_RES18","FPN_EF0"==> Outputs from the proposed ensemble can be generated using the Matlab script (commented out at the end of the file) provided by the authors.
-
Internal Testing:
- Requires the dataset folder to contain subdirectories named
JSRTandBSE_JSRT. - The
JSRTfolder should include the original X-ray images, whileBSE_JSRTshould contain the corresponding ground truth images.
- Requires the dataset folder to contain subdirectories named
-
External Testing:
- Automatically processes images in the specified folder, applies inference, and saves the output.
- Outputs are saved in the
outputsdirectory.
To train models, use the train.py script:
python train.py --model_name <desired_model> --data_path <dataset_directory>If you want to initialize the model with pre-trained weights, include the --weights_path argument:
python train.py --model_name <desired_model> --data_path <dataset_directory> --weights_path <weights_path>The dataset_directory should contain subdirectories named train and val. Both of these directories should contain JSRT and BSE_JSRT subdirectories with the same properties as described above.
This project is licensed under the MIT License - see the LICENSE file for details.
Schiller, V., Burget, R., Genzor, S., Mizera, J., & Mezina, A. (2025). xU-NetFullSharp: The Novel Deep Learning Architecture for Chest X-ray Bone Shadow Suppression. Biomedical Signal Processing and Control, 100, 106983. https://doi.org/10.1016/j.bspc.2024.106983. [link]
Bibtex:
@article{SCHILLER2025106983,
title = {xU-NetFullSharp: The Novel Deep Learning Architecture for Chest X-ray Bone Shadow Suppression},
journal = {Biomedical Signal Processing and Control},
volume = {100},
pages = {106983},
year = {2025},
issn = {1746-8094},
doi = {https://doi.org/10.1016/j.bspc.2024.106983},
url = {https://www.sciencedirect.com/science/article/pii/S1746809424010413},
author = {Vojtech Schiller and Radim Burget and Samuel Genzor and Jan Mizera and Anzhelika Mezina},
keywords = {Deep learning, Bone shadow suppression, X-ray images, Medical Imaging, Neural networks, Convolutional neural networks, U-Net, Image denoising},
}The xU-NetFullSharp is based on the most recent U-NetSharp [8] architecture and utilizes bidirectional multi-scale skip connections like in the preceding U-Net3+ [2]. The ReLU activation is changed for more modern xUnit [4] activation to ensure more accurate activation maps.
Blocks of the proposed architecture are made up of 2D convolutions with different dilation rates and xUnit [4] activation functions.
The experiments utilized three datasets – extensively augmented JSRT, VinDr-CXR [6], and Gusarev DES [1] dataset. The JSRT dataset, as well as the VinDr-CXR datasets, is available in the datasets folder in the cloud storage. The Gusarev DES dataset can be obtained from the following GitHub repository. Firstly, the JSRT dataset containing bone shadow-suppressed CXRs was split into training, validation, and testing sets and was extensively augmented to achieve a sufficient amount of usable images and to ensure the model’s robustness (both original and augmented images are available in the JSRT subfolder). The second, VinDr-CXR, dataset was augmented by randomly applying inversion and used for independent testing (the used testing set is available in the VinDrCXR subfolder). From the third, Gusarev DES, dataset expert pulmonologists selected images where the rib shadows collide with pulmonary nodules. These images were then used to conduct a performance assessment focused on clinical applications of the models.
The internal testing results (on the JSRT dataset) are available in the internal_test folder; external testing results (on the VinDr-CXR dataset) are present in the external_test folder. To reproduce the results, use the test.py file with the desired model and path to corresponding weights. Sample outputs from individual models can be seen in the /images/xray folder of this repository. The objective and subjective results we achieved on the individual datasets can be seen in the tables below.
| Models | MAE | MSE | SSIM | MS-SSIM | UIQI | PSNR [dB] |
|---|---|---|---|---|---|---|
| U-Net | 0.0074 | 0.0004 | 0.9835 | 0.9865 | 0.9959 | 34.6081 |
| Attention U-Net | 0.0080 | 0.0004 | 0.9834 | 0.9863 | 0.9960 | 34.4984 |
| Deep Residual U-Net | 0.0120 | 0.0007 | 0.9686 | 0.9768 | 0.9903 | 31.5990 |
| U-Net++ | 0.0073 | 0.0004 | 0.9834 | 0.9867 | 0.9959 | 34.6063 |
| Attention U-Net++ | 0.0076 | 0.0004 | 0.9835 | 0.9861 | 0.9957 | 34.4042 |
| U-Net3+ | 0.0074 | 0.0004 | 0.9836 | 0.9868 | 0.9957 | 34.6300 |
| U-NetSharp | 0.0078 | 0.0004 | 0.9836 | 0.9868 | 0.9960 | 34.4829 |
| xU-NetFullSharp | 0.0071 | 0.0003 | 0.9846 | 0.9870 | 0.9961 | 34.5338 |
| Attention xU-NetFullSharp | 0.0076 | 0.0003 | 0.9841 | 0.9869 | 0.9960 | 34.7016 |
| Kalisz Marczyk’s Autoencoder | 0.0097 | 0.0005 | 0.9722 | 0.9821 | 0.9948 | 33.3040 |
| FPN-ResNet-18 | 0.0154 | 0.0006 | 0.9708 | 0.9781 | 0.9917 | 31.7906 |
| FPN-EfficientNet-B0 | 0.0164 | 0.0006 | 0.9694 | 0.9802 | 0.9917 | 31.6370 |
| U-Net-ResNet-18 | 0.0172 | 0.0008 | 0.9660 | 0.9713 | 0.9904 | 30.7114 |
| DeBoNet | 0.0159 | 0.0022 | 0.9312 | 0.9642 | 0.9953 | 27.0403 |
| Models | Correlation | Intersection | 𝜒2 | Bhattacharyya |
|---|---|---|---|---|
| U-Net | 0.9443 | 9.8999 | 6.1702 | 0.1292 |
| Attention U-Net | 0.9449 | 9.8873 | 8.3925 | 0.1295 |
| Deep Residual U-Net | 0.9358 | 9.5917 | 5.8152 | 0.1478 |
| U-Net++ | 0.9394 | 9.8962 | 4.8495 | 0.1309 |
| Attention U-Net++ | 0.9478 | 9.9733 | 12.5015 | 0.1202 |
| U-Net3+ | 0.9329 | 9.9043 | 8.3654 | 0.1251 |
| U-NetSharp | 0.9414 | 9.9133 | 14.6634 | 0.1330 |
| Attention xU-NetFullSharp | 0.9321 | 9.8601 | 8.9438 | 0.1396 |
| xU-NetFullSharp | 0.9631 | 10.0285 | 7.0048 | 0.1155 |
| Kalisz Marczyk’s Autoencoder | 0.9580 | 9.8830 | 6.7579 | 0.1220 |
| FPN-ResNet-18 | 0.8658 | 9.1395 | 5.1873 | 0.1976 |
| FPN-EfficientNet-B0 | 0.8763 | 9.3075 | 12.9367 | 0.1831 |
| U-Net-ResNet-18 | 0.8843 | 9.0170 | 26.0124 | 0.1973 |
| DeBoNet | 0.9273 | 9.6682 | 11.0798 | 0.1418 |
| Models | Average rating (best = 1) |
Expert’s comment | |
| U-Net | 3.0 | Our models | |
| Attention U-Net | 2.8 | ||
| U-Net++ | 3.0 | ||
| Attention U-Net++ | 2.3 | ||
| U-Net3+ | 2.8 | Some problems with detail retention. | |
| U-NetSharp | 1.7 | Second best in bone shadow suppression. Great retention of details. | |
| xU-NetFullSharp | 1.2 | Consistently the best in bone shadow suppression. Great retention of details. | |
| Attention xU-NetFullSharp | 3.0 | ||
| U-Net-ResNet-18 | 4.5 | Concurrent models | |
| FPN-ResNet-18 | 4.3 | Major deformation of the chest silhouette in one sample! | |
| FPN-EfficientNet-B0 | 3.7 | Blurry details. | |
| Kalisz Marczyk’s Autoencoder | 3.5 |
| Models | Vessel visibility 3: Clearly visible 2: Visible 1: Not visible |
Airway visibility 3: Lobar and intermediate bronchi 2: Main bronchus and rump 1: Trachea |
Bone shadow suppression 3: Nearly perfect 2: Less than 5 unsuppressed bones 1: 5 or more unsuppressed bones |
Overall bone shadow suppression performance 1: Excellent 3: Average 5: Poor |
Nodule visibility 3: More apparent 2: Equally as apparent 1: Less apparent |
| DES (Reference) | 3 | 1.8 | 2.7 | 1.3 | 2.9 |
| Attention U-Net | 3 | 2.1 | 1 | 2.8 | 2.3 |
| Attention U-Net++ | 3 | 2.2 | 1 | 2.3 | 2.1 |
| Attention xU-NetFullSharp | 3 | 2.3 | 1 | 2.4 | 2 |
| DeBoNet | 2 | 1.8 | 1 | 4.2 | 1.4 |
| FPN-EfficientNet-B0 | 3 | 2.2 | 1 | 3.3 | 1.5 |
| FPN-ResNet-18 | 3 | 2.3 | 1 | 3.5 | 2 |
| U-Net-ResNet-18 | 3 | 2.3 | 1 | 4.9 | 2.1 |
| Deep Residual U-Net | 2.2 | 2 | 1 | 5 | 1.9 |
| Kalisz Marczyk’s Autoencoder | 3 | 2.3 | 1 | 3.1 | 2.1 |
| U-Net | 3 | 2.3 | 1 | 2.6 | 2.2 |
| U-NetSharp | 3 | 2.4 | 1 | 2.1 | 2.2 |
| U-Net3+ | 3 | 2.4 | 1 | 2.5 | 2.2 |
| U-Net++ | 3 | 2.4 | 1 | 2.2 | 2.1 |
| xU-NetFullSharp | 3 | 2.4 | 1 | 1.6 | 2.3 |
[1] Gusarev, M., Kuleev, R., Khan, A., Ramirez Rivera, A., & Khattak, A. M. (2017). Deep learning models for bone suppression in chest radiographs. 2017 IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB), 1–7. https://doi.org/10.1109/CIBCB.2017.8058543
[2] Huang, H., Lin, L., Tong, R., Hu, H., Zhang, Q., Iwamoto, Y., Han, X., Chen, Y.-W., & Wu, J. (2020). UNet 3+: A Full-Scale Connected UNet for Medical Image Segmentation. ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1055–1059. https://doi.org/10.1109/ICASSP40776.2020.9053405
[3] Kalisz, S., & Marczyk, M. (2021). Autoencoder-based bone removal algorithm from x-ray images of the lung. 2021 IEEE 21st International Conference on Bioinformatics and Bioengineering (BIBE), 1–6.
[4] Kligvasser, I., Shaham, T. R., & Michaeli, T. (2017). xUnit: Learning a Spatial Activation Function for Efficient Image Restoration. CoRR, abs/1711.06445. http://arxiv.org/abs/1711.06445
[5] Li, C., Tan, Y., Chen, W., Luo, X., Gao, Y., Jia, X., & Wang, Z. (2020). Attention Unet++: A Nested Attention-Aware U-Net for Liver CT Image Segmentation. 2020 IEEE International Conference on Image Processing (ICIP), 345–349. https://doi.org/10.1109/ICIP40778.2020.9190761
[6] Nguyen, H. Q., Lam, K., Le, L. T., Pham, H. H., Tran, D. Q., Nguyen, D. B., Le, D. D., Pham, C. M., Tong, H. T. T., Dinh, D. H., Do, C. D., Doan, L. T., Nguyen, C. N., Nguyen, B. T., Nguyen, Q. V., Hoang, A. D., Phan, H. N., Nguyen, A. T., Ho, P. H., … Vu, V. (2022). VinDr-CXR: An open dataset of chest X-rays with radiologist’s annotations.
[7] Oktay, O., Schlemper, J., Folgoc, L. le, Lee, M. C. H., Heinrich, M. P., Misawa, K., Mori, K., McDonagh, S. G., Hammerla, N. Y., Kainz, B., Glocker, B., & Rueckert, D. (2018). Attention U-Net: Learning Where to Look for the Pancreas. CoRR, abs/1804.03999. http://arxiv.org/abs/1804.03999
[8] Qian, L., Zhou, X., Li, Y., & Hu, Z. (2022). UNet#: A UNet-like Redesigning Skip Connections for Medical Image Segmentation. ArXiv Preprint, ArXiv:2205.11759.
[9] Rajaraman, S., Cohen, G., Spear, L., Folio, L., & Antani, S. (2022). DeBoNet: A deep bone suppression model ensemble to improve disease detection in chest radiographs. Plos One, 17(3), e0265691.
[10] Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation.
[11] Zhang, Z., Liu, Q., & Wang, Y. (2017). Road Extraction by Deep Residual U-Net. CoRR, abs/1711.10684. http://arxiv.org/abs/1711.10684
[12] Zhou, Z., Siddiquee, M. M. R., Tajbakhsh, N., & Liang, J. (2018). UNet++: A Nested U-Net Architecture for Medical Image Segmentation. CoRR, abs/1807.10165. http://arxiv.org/abs/1807.10165