This project used to segment the LiveCell dataset

git clone https://github.com/rishabhmohatta/LiveCell_segmentation.git
cd LiveCell_segmentation

pip install -r requirements.txt

#Downlad the weights from this link and copy in folder LiveCell_segmentation
https://drive.google.com/file/d/1vj3glsdqkCFdSJrTMgESszeMl2LhRd5S/view?usp=sharing

uvicorn app:app --host 0.0.0.0 --port 8000

curl -X POST -F "file=@test.jpg" "http://127.0.0.1:8000/predict/" --output result.png

docker build -t unet-api .

docker run -p 8000:8000 unet-api

The U-Net model was trained on the LIVECell dataset using:

• Model: Unet
• Input Image Size: 256x256

To evaluate the model performance, mAP (Mean Average Precision) was calculated on the test dataset.

python evaluate.py --model unet_livecell_best.pth --images test/ --labels test.json [--num_images]

Example output:

For 100 test images: mAP: 0.84, IoU: 0.81
For total(1500) test images : mAP: 0.83, IoU: 0.78

• U-Net Architecture Paper: Ronneberger et al., 2015
• LIVECell Dataset Paper: Edlund et al., 2021
• PyTorch U-Net Implementation: GitHub - milesial/Pytorch-UNet

• Model trained on single class as Dataset has 8 type of cell can create a multiclass model
• Optimizing model for real time inference (like converting it into onnx ,trt,etc)
• Training the model with an pretrained resnet architecture as a backbone for better accuracy
• Using other different model like deeplab for segmentation