| <h1 style="border-bottom: 2px solid black; font-size: 100px;" align="center"> SwinUNETR </h1> | |
| _Trained by Margerie Huet Dastarac ._ <br> | |
| _Training date: November 2023 ._ | |
| ## 1. Task Description | |
| Segmentation of the body on the CT scan on a dataset of 60 oropharyngeal patients. This model can be used to clean CT scans by setting voxels value outside of the body contour to air, a typical preprocessing step for other networks. | |
| ## 2. Model | |
| ### 2.1. Architecture | |
| <img width="100%" src="https://cdn-uploads.huggingface.co/production/uploads/65c9dbefd6cbf9dfed67367e/o59In69BqrxTEoOdSHZLD.png" alt="alternatetext"> | |
| _Figure 1: SwinUNETR architecture_ | |
| ### 2.2. Input | |
| <ul> | |
| <li> CT</li> | |
| </ul> | |
| ### 2.3. Output | |
| <ul> | |
| <li> BODY</li> | |
| </ul> | |
| ### 2.4 Training details | |
| <ul> | |
| <li> Number of epoch: 300 </li> | |
| <li> Loss function: Dice loss </li> | |
| <li> Optimizer: Adam </li> | |
| <li> Learning Rate: 3e-4 </li> | |
| <li> Dropout: No </li> | |
| <li> Patch size in voxels: (128,128,128) </li> | |
| <li> Data augmentation used: | |
| <ul> | |
| <li> RandSpatialCropd</li> | |
| <li> RandFlipd axis:0</li> | |
| <li> RandFlipd axis:1</li> | |
| <li> RandFlipd axis:2</li> | |
| <li> NormalizeIntensityd</li> | |
| <li> RandScaleIntensityd factors:0.1 prob:1.0</li> | |
| <li> RandShiftIntensityd, offsets:0.1, prob:1.0</li> | |
| </ul> | |
| </li> | |
| </ul> | |
| ## 3. Dataset | |
| <ul> | |
| <li> Location: Head and neck, oropharynx </li> | |
| <li> Training set size: 60 </li> | |
| <li> Resolution in mm: 3x3x3 </li> | |
| </ul> | |
| ## Performance | |
| + TBD | |