Implementation

• The official code of TransUNet is used for the model 2D implementation. Intermediate outputs are added in the model to perform deep supervision training.
• For TransUNet3D, all convolution are changed to Conv3D.
• To train with nnUNet, a planner and trainer are also implemented:
  • Planner:
    • Determine the batch size based on GPU maximum memory
    • Define the image patch size
  • Trainer: inspired from nnUNetTrainerV2, this trainer is adapted to instantiate the model, to calculate deep supervision loss weights based on TransUNet configuration.

Training

• Train 1 fold:

nnUNet_train MODEL TransUNetTrainer TASK_NAME FOLD

MODEL could be TransUNet2D or TransUNet3D

• After training all 5 folds, we need to determine the postprocessing of ensemble of these 5 folds:

nnUNet_find_best_configuration -m TransUNet3D -t TASK_NUMBER

Inference

nnUNet_predict -i INPUT_IMAGE_DIR -o OUTPUT_INFERENCE_DIR -m TransUNet3D --save_npz -t TASK_NAME

Evaluate

nnUNet_evaluate_folder -ref TARGET_IMAGE_DIR -pred PREDICTION_DIR -l L [Class number like 1 2 3 4 5]