MedficientSAM: A Robust Medical Segmentation Model with Optimized Inference Pipeline for Limited Clinical Settings

Getting Started

git clone --recursive https://github.com/hieplpvip/medficientsam.git
cd medficientsam
conda env create -f environment.yaml -n medficientsam
conda activate medficientsam

Environment and Requirements

System	Ubuntu 22.04.5 LTS
CPU	AMD EPYC 7742 64-Core Processor
RAM	256GB
GPU (number and type)	One NVIDIA A100 40G
CUDA version	12.0
Programming language	Python 3.10
Deep learning framework	torch 2.2.2, torchvision 0.17.2

Results

Accuracy metrics are evaluated on the public validation set of CVPR 2024 Segment Anything In Medical Images On Laptop Challenge. The computational metrics are obtained on an Intel(R) Core(TM) i9-10900K.

Model	Resolution	Params	FLOPs	DSC	NSD	2D Run Time	3D Run Time	2D Memory Usage	2D Memory Usage
MedficientSAM-L0	512x512	34.79M	36.80G	85.85%	87.05%	0.9s	7.4s	448MB	687MB
MedficientSAM-L1	512x512	47.65M	51.05G	86.42%	87.95%	1.0s	9.0s	553MB	793MB
MedficientSAM-L2	512x512	61.33M	70.71G	86.08%	87.53%	1.1s	11.1s	663MB	903MB

Reproducibility

The Docker images can be found here.

docker load -i seno.tar.gz
docker container run -m 8G --name seno --rm -v $PWD/test_input/:/workspace/inputs/ -v $PWD/test_output/:/workspace/outputs/ seno:latest /bin/bash -c "sh predict.sh"

To measure the running time (including Docker starting time), see https://github.com/bowang-lab/MedSAM/blob/LiteMedSAM/CVPR24_time_eval.py

Data Preparation

You need to participate in the challenge to access the dataset. After downloading it, copy .env.example to .env and modify CVPR2024_MEDSAM_DATA_DIR to correct path.

The directory structure should look like this:

CVPR24-MedSAMLaptopData
├── train_npz
│   ├── CT
│   ├── Dermoscopy
│   ├── Endoscopy
│   ├── Fundus
│   ├── Mammography
│   ├── Microscopy
│   ├── MR
│   ├── OCT
│   ├── PET
│   ├── US
│   └── XRay
├── validation-box
    └── imgs

Training

See train_scripts.

Inference

You can download the weights here.

To run MedficientSAM-L1 on the validation set:

python src/infer.py experiment=infer_finetuned_l1

See configs/experiment/infer_* for running other model variants.

Build Docker image

Export model

python src/export_onnx.py experiment=export_finetuned_l0_cpp output_dir=weights/finetuned-l1-augmented/e2_cpp

Build

docker build -f Dockerfile.cpp -t seno.fat .
slim build --target seno.fat --tag seno --http-probe=false --include-workdir --mount $PWD/test_input/:/workspace/inputs/ --mount $PWD/test_output/:/workspace/outputs/ --exec "sh predict.sh"
docker save seno | gzip -c > seno.tar.gz

hieplpvip/medficientsam