The SAM trained on 2D natural images used by nnsam has limited improvement for 3D medical images. Therefore, using Med SAM, which is pre-trained on medical image datasets, is better suited to the task of segmenting medical images.

checkpoint for Med SAM 3D: SAM-Med3D-Turbo

Our entire code is built based on nnUNet, and you can follow the nnUNet instructions exactly.

conda create -n nnsam python=3.9
conda activate nnsam
pip install git+https://github.com/ChaoningZhang/MobileSAM.git
pip install timm
pip install git+https://github.com/SuperJunier666/nnMed-SAM.git

It is important to input "set MODEL_NAME=nnsam" before using it.

set MODEL_NAME=nnsam

nnUNetv2_plan_and_preprocess -d DATASET_ID --verify_dataset_integrity

nnUNetv2_train DATASET_NAME_OR_ID UNET_CONFIGURATION FOLD [additional options, see -h]

nnUNetv2_train DATASET_NAME_OR_ID UNET_CONFIGURATION FOLD --val --npz

nnUNetv2_train DATASET_NAME_OR_ID 2d FOLD

nnUNetv2_train DATASET_NAME_OR_ID 3d_fullres FOLD

nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_NAME_OR_ID -c CONFIGURATION --save_probabilities

How to get started?

Read these:

• Dataset conversion
• Usage instructions

Additional information:

• Region-based training
• Manual data splits
• Pretraining and finetuning
• Intensity Normalization in nnU-Net
• Manually editing nnU-Net configurations
• Extending nnU-Net
• What is different in V2?