Skip to content

Latest commit

 

History

History
68 lines (47 loc) · 2.82 KB

INSTRUCTIONS.md

File metadata and controls

68 lines (47 loc) · 2.82 KB

Requirements

CUDA version: 11.8.

torch==1.13.0 ; torchaudio==0.13.0 ; torchvision==0.14.0.

You can check the propper torch version for your CUDA at: https://pytorch.org/get-started/previous-versions/

Data acquisition

Download any BraTS dataset from Synapse following their instructions.

Link: https://www.synapse.org/Synapse:syn51156910/wiki/627000

The dataset folder structure should be as follows:

-data
--brats #Adult Glioma Segmentation
--brats_ssa #Subsaharan Glioma Segmentation
--brats_ped #Pediatric Glioma Segmentation
--brats_men #Meningioma Segmentation

You can download as much BraTS datasets as you want to use, but all of them should be placed inside the data folder following the above specified naming instructions.

Download code & Set the environment

Open a terminal and execute the following commands:

git clone https://github.com/vpulab/med-sam-brain/;
cd med-sam-brain;
conda env create -f environment.yml;
conda activate sam_adapt_brain;

Training & Testing

Training

python train.py -net sam -mod sam_lora -exp_name ... -sam_ckpt ./checkpoint/sam/sam_vit_b_01ec64.pth -b 1 -dataset brats -thd True  -data_path ../data -w 8 -four_chan True
  • mod can be defined as: sam_lora to train LoRA blocks making SAM adapt to the medical domain; or sam in case you want to maintain the original SAM architecture.
  • four_chan should be defined as True if you want to use all 4 MRI modalities; or False if just taking e of them to not train the Patch Embedding Layer.
  • dataset must be defined as any of the names indicated in the 'Data acquisition' section.
  • exp_name should be replaced by the desired name of the experiment.

NOTE: After running the training command, 'sam_vit_b_01ec64.pth' will be downloaded. If pretrained weights are not downloaded propperly, you cand do it manually through this link and store it in 'checkpoint/sam/'. The saved model parameters will be placed in the 'logs/' directory.

Validation

python val.py -net sam -mod sam_lora -thd True  -dataset brats -weights logs/.../Model/best_dice -sam_ckpt logs/.../Model/best_dice -mode Validation -four_chan True
  • weights and sam_ckpt should be replaced by the directory of the saved model file in 'logs/'.
  • exp_nameshould be replaced by the training experiment name.

NOTE: In case you don't have enough GPU to execute the training process, you can uncomment the following code lines on function.py, which reduces computational cost by taking 4 random slices per volume (the selected slices change each iteration).

# If not enough GPU, uncomment the following 3 lines (lines 73-76 and 226-230)
# i_slices = SelectEquiSlices(4, masks)
# imgs = imgs[:,:,:,:,i_slices] 
# masks = masks[:,:,:,:,i_slices]