UC3 UNITOBrain

This repository contains the source code for getting started on UniToBrain dataset by using pyEDDL/pyECVL

Requirements

numpy
pyeddl
pyecvl
cv2
wandb
scipy
pydicom

1) Preprocessing Run Command

The network inputs are tensors of multiple CT images at the same scansion height for each patient. This script creates a folder called input_tensored into the prep_output_path path.

Image size of 128 pixels is used for the pratrain phase, size of 512 for the full resolution training

# 1.1) compute the inputs
python3 -u dicomtopt.py --rescale_size 128 --prep_output_path <output-path> --unitobrain_path <unitobrain-path> --target 'INPUT'

# 1.2) compute the target perfusion maps
python3 -u dicomtopt.py --rescale_size 128 --prep_output_path <output-path> --unitobrain_path <unitobrain-path> --target 'TTP'
python3 -u dicomtopt.py --rescale_size 128 --prep_output_path <output-path> --unitobrain_path <unitobrain-path> --target 'CBF'
python3 -u dicomtopt.py --rescale_size 128 --prep_output_path <output-path> --unitobrain_path <unitobrain-path> --target 'CBV'

2) Pretrain Run Command (Optional)

Pretrained model on lower resolution tensors (target TTP on 4 gpus)

python3 -u train_model.py --target 'TTP' --shape 128 --lr 1e-5 --num_gpu 4 --epochs 100 --batch-size 8 --mem 'low_mem' --name <run-name> <prep_output_path>

3) Train Run Command

Train the model on full resolution tensors (target TTP on 4 gpus)

python3 -u train_model.py --target 'TTP' --resume_ckpts <pretrain-checkpoint> --batch-size 8 --batch-size-val 4 --lr 1e-5 --epochs 50 --num_gpu 4 --name <run-name> --mem 'low_mem' --shape 512 --log-interval 20 <prep_output_path>

4) Inference Run Command

Inference command to run Tests only