Code for:

A Comparison of Deep Learning Methods for Cell Detection in Digital Cytology

M. Acerbis, N. Sladoje, and J. Lindblad

ArXiv pre-print SCIA2025

Dataset

The CNSeg Dataset is made publicly available on Kaggle.

Requisites

pip install requisites.txt

Segmentation-based Methods

• StarDist: modify the requested paths to data and labels. Then run:

python stardist_test.py

• Cellpose: modify the requested paths to data and labels. Then run:

python cellpose_test.py

• SAM2: pre-trained SAM2 models can be found here. Modify the requested paths to data and labels. Then run:

python SAM2_test.py

Centroid-based Methods

• FCRN - IFCRN In the respective folders, you can find the code to:
  • Build each model architecture,
  • Train-from-scratch
    FRCN: modify the requested paths and then run

    python fcrn_train.py
    

    and IFCRN: run the following

    python train_cnseg.py --train-img-dir --train-mask-dir --val-img-dir --val-mask-dir -e -b 
    

  • Evalaute the models
    FRCN: modify the requested paths and then run

    python fcrn_test.py
    

    and IFCRN: run the following

    python predict.py -i -o -m -t --labels-dir
    

Note that path to data, labels, and binary masks (for training) have to be passed to the parser.