pytorch-upper-Maxilla-and-lower-Mandible-jaws-segmentation

pytorch segmentation for upper (Maxilla) and lower(Mandible) jaws in a DICOM file. The Dataset is provided as 2D slices from all 3 orthogonal points of view Axial, Coronal, and Sagittal.

Test predictions Sample

---

Download open source dataset and explore it!

Download_jaw_segmentation_DS.ipynb contains the required code for download and inspect the dataset.

---

Install Requirments python >= 3.8

python3 -m pip install -r requirements.txt

---

Repo Structure

All the code for training and testing found under source/.

.
├── Download_jaw_segmentation_DS.ipynb (Download and explore the dataset)
├── LICENSE
├── README.md
├── requirements.txt
├── source
│   ├── baseline.py (entry for a single experiment run)
│   ├── experiment_handler.py (load multiple experiments configs from experiments.json)
│   ├── experiments.json (contains configerations for list of experiement that can be run sequentially)
│   ├── helpers
│   │   └── helpers.py (contains helper function used for logging and plotting, etc ...)
│   ├── learning
│   │   ├── dataset.py (implement the dataset loaders)
│   │   ├── learner.py (implement the training, eval, and test functions)
│   │   ├── losses.py (implement multiple loss function)
│   │   ├── model.py (implment multiple semantic seg models, currently UNET)
│   │   └── utils.py (contains all the utils, reading the configs and get dataset transform functions, get model, etc.)
│   └── option.py (contains arguments config if you want to pass them through the terminal instead of a json file)
└── view
└── Dataset.png

---

Training and Baseline

• using experiment.json

python3 baseline.py --save_path path/to/save/exp/output --model model_name --train_size 128 128 --test_size 128 128 --crop_size --batch_size 8; # other parameters found in [options.py](source/option.py)

• using terminal
  • set the configurations of the experiements in experiments.json, there are 2 main keys, the general config which have general configurations for all experiments, and experiments which contains list of experiments config which overwrites the general config based on each experiments customizatiom.

python3 source/experiment_handler.py # after set the configurations

---

Supported Losses

• Cross Entropy
• Dice Loss
• Focal Loss
• Travesky Loss

---

Supported Models

• UNET

---

Reference

• Focal Loss
• Class Weighted CE Loss
• EvoNorm
• StarterCode-@hoya012