Segmentation example

The segmentation example is the smallest, most conservative training workflow shipped with KonfAI. It is meant to be easy to read, easy to adapt, and easy to use as a first project template.

What is in the folder

examples/Segmentation/
├── Config.yml
├── Prediction.yml
├── Evaluation.yml
└── Segmentation_demo.ipynb

  • Config.yml defines the training workflow.

  • Prediction.yml defines inference and export.

  • Evaluation.yml computes Dice on the saved predictions.

  • Segmentation_demo.ipynb bootstraps the example in a fresh environment.

Expected dataset layout

Dataset/
├── CASE_000/
│   ├── CT.mha
│   └── SEG.mha
└── ...

  • CT is the input image.

  • SEG is the segmentation label map.

The shipped demo assumes:

  • a multiclass task

  • label 0 for background

  • labels 1..40 for foreground classes

  • 41 classes total

Default baseline

The baseline uses:

  • the built-in segmentation.UNet.UNet

  • 2D patch-based training

  • CrossEntropyLoss during training

  • Dice during evaluation

Training uses CrossEntropyLoss on purpose. It keeps the live training loop simple and stable while final quality is still measured with Dice during evaluation.

Minimal workflow

Run all commands from examples/Segmentation.

Train:

konfai TRAIN -y --gpu 0 --config Config.yml

Predict:

konfai PREDICTION -y --gpu 0 --config Prediction.yml \
  --models Checkpoints/SEG_BASELINE/<checkpoint>.pt

Evaluate:

konfai EVALUATION -y --config Evaluation.yml

Outputs are written to:

  • Checkpoints/SEG_BASELINE/

  • Statistics/SEG_BASELINE/

  • Predictions/SEG_BASELINE/

  • Evaluations/SEG_BASELINE/

What to adapt first

Most real projects will need changes in:

  1. dataset_filenames

  2. train_name

  3. patch size

  4. batch size

  5. nb_class

  6. preprocessing transforms

  7. the list of Dice labels in Evaluation.yml

If your dataset is not a 0..40 label map, update both nb_class and the Dice labels together.

See also