Awesome

[Swin-LiteMedSAM] Submission to the SEGMENT ANYTHING IN MEDICAL IMAGES ON LAPTOP 2024 CVPR challenge [Teamname: lkeb]

Install

• Create a virtual environment conda create -n swin_litemedsam python=3.10 -y and activate it conda activate swin_litemedsam
• Install Pytorch 2.x
• Enter the Swin_MedSAM folder cd Swin_LiteMedSAM and run pip install -e .

Model

• Download the model checkpoint and place it at workdir
• Donwload the docker image and load it docker load -i lkeb.tar.gz for quick inference

Usage

1. Data

   Download training npz data from challenge website, the training data contained 11 modalities, including CT, MRI, PET, X-Ray, ultrasound, mammography, OCT, endoscopy, fundus, dermoscopy, and microscopy.

2. Train

   • Distill encoder

   cd distill
   python train_distill.py
   

   • Only train decoder

   python train.py
   

   • Train encoder & decoder together

   python train.py -freeze False
   

3. Docker Infer

   • Load the docker

   docker load -i lkeb.tar.gz
   

   • Run the docker infer case in $PWD/imgs

   docker container run -m 8G --name lkeb --rm -v $PWD/imgs/:/workspace/inputs/ -v $PWD/lkeb/:/workspace/outputs/ lkeb:latest /bin/bash -c "sh predict.sh"
   

   • if you run into Permission denied error.

   chmod -R 777 ./*