High-quality tomographic image reconstruction integrating neural networks and mathematical optimization
Fuente:
Zenodo
Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Recurso digital |
| Langue: | anglais |
| Publié: |
Zenodo
2025
|
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866902332889890816 |
|---|---|
| author | Mishra, Anuraag |
| author_facet | Mishra, Anuraag |
| contents | <p>This Zenodo record accompanies the article “High-quality tomographic image reconstruction integrating neural networks and mathematical optimization” published in Machine Learning: Science and Technology, 2025, <a title="">DOI 10.1088/2632-2153/ae25b6</a>. It contains the source code, trained neural network files, serialized training data, example images, and environment specifications used to study hybrid tomographic reconstruction methods that combine deep learning with mixed-integer optimization.</p> <p>The archive is organized around two connected workflows. The first workflow generates local 3x3 grayscale image patches from an example image, computes Sobel-based edge magnitudes, and trains a compact fully connected neural network to predict local edge intensity. The second workflow integrates the trained network into two optimization-based reconstruction approaches built on the CSHM framework for homogeneous materials: direct DNN-to-MIP integration and DNN-to-MIP re-optimization. These methods are designed to improve edge sharpness and material homogeneity in reconstructions obtained from sparse or inconsistent projection data.</p> <p>The deposited files include Python source code, notebooks, Conda environment files, PNG example images, trained PyTorch checkpoints, and auxiliary binary files required by the reconstruction pipeline. In particular, PKL files store serialized pandas DataFrames or exported Python and NumPy objects such as training patches, Sobel target values, and learned network weights and biases. PTH files store PyTorch model checkpoints. These binary files are machine-readable research objects and are intended to be loaded with Python libraries such as pandas, pickle, NumPy, and PyTorch rather than opened as plain-text files.</p> <p>The methods were evaluated on simulated phantom data and on experimental tomography examples, including electron tomography of a zeolite particle and nano-CT of a copper microlattice. This record is provided to document the computational workflow behind the publication, support reproducibility, and enable reuse of the DNN training and optimization pipeline for related reconstruction studies. Large raw projection datasets are not fully redistributed in all cases; the included README explains the archive contents, file formats, and the role of binary files, while the code documents the supported projection formats and expected directory structure for external tomography data.</p> <p><strong>Keywords</strong></p> <p>tomographic image reconstruction</p> <p>electron tomography</p> <p>nano-CT</p> <p>deep neural networks</p> <p>mixed-integer linear programming</p> <p>compressed sensing</p> <p>edge detection</p> <p>hybrid optimization</p> <p>zeolite</p> <p>copper microlattice</p> <p>homogeneous materials</p> <p>Sobel operator</p> <p><strong>Funding</strong></p> <p>Deutsche Forschungsgemeinschaft, DFG, German Research Foundation, Project-ID 416229255, CRC 1411</p> <p>Related project link: <a title="">CRC 1411 project website</a></p> <p><strong>Related identifier</strong></p> <p>Associated publication: <a title="">DOI 10.1088/2632-2153/ae25b6</a></p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18909837 |
| institution | Zenodo |
| language | eng |
| publishDate | 2025 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | High-quality tomographic image reconstruction integrating neural networks and mathematical optimization Mishra, Anuraag <p>This Zenodo record accompanies the article “High-quality tomographic image reconstruction integrating neural networks and mathematical optimization” published in Machine Learning: Science and Technology, 2025, <a title="">DOI 10.1088/2632-2153/ae25b6</a>. It contains the source code, trained neural network files, serialized training data, example images, and environment specifications used to study hybrid tomographic reconstruction methods that combine deep learning with mixed-integer optimization.</p> <p>The archive is organized around two connected workflows. The first workflow generates local 3x3 grayscale image patches from an example image, computes Sobel-based edge magnitudes, and trains a compact fully connected neural network to predict local edge intensity. The second workflow integrates the trained network into two optimization-based reconstruction approaches built on the CSHM framework for homogeneous materials: direct DNN-to-MIP integration and DNN-to-MIP re-optimization. These methods are designed to improve edge sharpness and material homogeneity in reconstructions obtained from sparse or inconsistent projection data.</p> <p>The deposited files include Python source code, notebooks, Conda environment files, PNG example images, trained PyTorch checkpoints, and auxiliary binary files required by the reconstruction pipeline. In particular, PKL files store serialized pandas DataFrames or exported Python and NumPy objects such as training patches, Sobel target values, and learned network weights and biases. PTH files store PyTorch model checkpoints. These binary files are machine-readable research objects and are intended to be loaded with Python libraries such as pandas, pickle, NumPy, and PyTorch rather than opened as plain-text files.</p> <p>The methods were evaluated on simulated phantom data and on experimental tomography examples, including electron tomography of a zeolite particle and nano-CT of a copper microlattice. This record is provided to document the computational workflow behind the publication, support reproducibility, and enable reuse of the DNN training and optimization pipeline for related reconstruction studies. Large raw projection datasets are not fully redistributed in all cases; the included README explains the archive contents, file formats, and the role of binary files, while the code documents the supported projection formats and expected directory structure for external tomography data.</p> <p><strong>Keywords</strong></p> <p>tomographic image reconstruction</p> <p>electron tomography</p> <p>nano-CT</p> <p>deep neural networks</p> <p>mixed-integer linear programming</p> <p>compressed sensing</p> <p>edge detection</p> <p>hybrid optimization</p> <p>zeolite</p> <p>copper microlattice</p> <p>homogeneous materials</p> <p>Sobel operator</p> <p><strong>Funding</strong></p> <p>Deutsche Forschungsgemeinschaft, DFG, German Research Foundation, Project-ID 416229255, CRC 1411</p> <p>Related project link: <a title="">CRC 1411 project website</a></p> <p><strong>Related identifier</strong></p> <p>Associated publication: <a title="">DOI 10.1088/2632-2153/ae25b6</a></p> |
| title | High-quality tomographic image reconstruction integrating neural networks and mathematical optimization |
| url | https://doi.org/10.5281/zenodo.18909837 |