MobileSAM + LoRA for Nuclei Instance Segmentation on NuInsSeg

This repository trains a LoRA-based MobileSAM model for nuclei instance segmentation on the NuInsSeg dataset. It uses ground-truth bounding-box prompts derived from instance annotations and evaluates performance with five-fold cross-validation using Dice, AJI, and PQ.

download the NuInsSeg dataset via the shell script

chmod +x download_data.sh
./download_data.sh

install the base conda env, then add mobileSAM and other dependencies

conda env create -f env.yml
conda activate ml_med_imaging 
pip install tifffile opencv-python scikit-image scikit-learn timm
pip install git+https://github.com/ChaoningZhang/MobileSAM.git

download mobilesam weights

mkdir -p checkpoints
wget -O checkpoints/mobile_sam.pt https://github.com/ChaoningZhang/MobileSAM/raw/master/weights/mobile_sam.pt

run with

conda activate ml_med_imaging
python train_mobilesam_lora_nuinsseg.py