CVPR24-MedSAM-on-Laptop

This repository contains the code for our submission to the CVPR 2024: SEGMENT ANYTHING IN MEDICAL IMAGES ON LAPTOP competition:

DAFT: Data-Aware Fine-Tuning of Foundation Models for Efficient and Effective Medical Image Segmentation

BibTex citation: @InProceedings{pfefferle2025daft, author={Pfefferle, Alexander and Purucker, Lennart and Hutter, Frank}, editor={Ma, Jun and Zhou, Yuyin and Wang, Bo}, title={DAFT: Data-Aware Fine-Tuning of Foundation Models for Efficient and Effective Medical Image Segmentation}, booktitle={Medical Image Segmentation Foundation Models. CVPR 2024 Challenge: Segment Anything in Medical Images on Laptop}, year={2025}, publisher={Springer Nature Switzerland}, address={Cham}, pages={15--38}, isbn={978-3-031-81854-7}, doi={10.1007/978-3-031-81854-7_2} }

The competition was organized by BoWang's Lab at University of Toronto and hosted on Codabench.

The main branch contains the code of the performance booster challenge that followed the competition. Check out the finalsubmission branch for our final submission to the competition.

Environments and Requirements

We trained on the JUWELS Booster (4x A100 40GB, Python 3.11.3, CUDA 12.2) using the packages in requirements.txt.

Dataset and Models

• download the LiteMedSAM weights and put them in work_dir/LiteMedSAM/, also download the EfficientViT-SAM l0 checkpoint
• download the training data (we only used the data provided by the organizers, no other external allowed datasets)
• download the test data to CVPR24-MedSAMLaptopData and unzip it
• cd train_npz and unzip all files
• prepare the new datasets
  • mv NewTrainingData/totalseg_mr MR
  • cd CVPR24-PostChallenge-Train/ && mkdir XRay PET
  • for file in *; do [ -f "$file" ] && mv "$file" "${file:3}"; done (we want all files of a modality to have the modality name as prefix)
  • mv X-Ray_Teeth_* XRay/ && mv PET_Lesion_psma_* PET/

Now we just need to split the data into training and validation sets and foundation modalities for DAFT: + go to wherever you cloned the repository and run python split_dataset.py <path_to_data> + cd datasplit && python modalities3D.py train.csv val.csv

Preprocessing

• we didn't do any further preprocessing outside the distillation/finetuning code
• during distillation/finetuning we did the same as the baseline training code
  • resize and pad images to 256x256
  • normalize intensities
  • flip horizontally with 50% probability
  • flip vertically with 50% probability
  • randomly pertubate the box by up to 5 pixels

Training

Make sure to use different workdirs for distillation, general finetuning and each DAFT run.

1. use python distill.py -num_epochs 70 -batch_size 7 -device cuda -work_dir work_dir_distill/ -resume work_dir_distill/medsam_lite_latest.pth -pretrained_checkpoint l0.pt to distill the TinyViT in LiteMedSAM to EfficientViT, use python modelmerge.py work_dir_distill/medsam_lite_best.pth distilled.pth to create weights for EfficientViT-SAM by using the distilled EfficientViT image encoder checkpoint and copying the prompt encoder and mask decoder from LiteMedSAM
2. fine tune our the model from step 3 on all the data by using python finetune.py -pretrained_checkpoint distilled.pth -num_epochs 70 -batch_size 96 -device cuda -work_dir work_dir_general -resume work_dir_general/medsam_lite_latest.pth, use python extract_evit.py work_dir_general/medsam_lite_best.pth general_finetuned.pth to extract the weights from the latest training checkpoint afterwards
3. fine tune our model from step 2 on different subsets of the data, depending on the modality, by using python daft.py -pretrained_checkpoint general_finetuned.pth -num_epochs 70 -batch_size 96 -device cuda -work_dir work_dir_modalities3D/<modality>/ -resume work_dir_modalities3D/<modality>/medsam_lite_latest.pth --traincsv datasplit/modalities3D/<modality>.train.csv --valcsv datasplit/modalities3D/<modality>.val.csv, afterwards use mkdir models && ./extract_modalities3D.sh to extract the weights from the checkpoints
4. cp general_finetuned.pth models/general.pth to use the generally finetuned model as a fallback if we can't read the modality from the file name
5. convert all PyTorch models to ONNX by running python export_onnx.py
6. cd onnxmodels && mv 3D_prompt_encoder.onnx prompt_encoder.onnx && rm *_prompt_encoder.onnx && cd .. since the prompt encoder is shared
7. convert ONNX to OpenVINO IR via python onnx2openvino.py

You can now use PerfectMetaOpenVINO.py for inference.

The OpenVINO IR artifacts of our performance booster submission and the distilled and generally finetuned models are also available here.

Inference

Download the demo data and run python PerfectMetaOpenVINO.py

Our C++ implementation is available in cpp/.

Evaluation

Run python evaluation/compute_metrics.py -s test_demo/segs -g test_demo/gts -csv_dir ./metrics.csv

Docker submission

To create the docker you can use ./build_docker_cpp.sh (or ./build_docker_openvino.sh for the python version) and docker save automlfreiburg | gzip -c > automlfreiburg.tar.gz to save it.

Meta Model Experiments

• run python metadataset.py to create the meta dataset
• run python metaag.py to train autogluon on the meta dataset and evaluate it on the test set
• run python distributionshift.py to check for a distribution shift between the training and test set