This repository hosts the code necessary to replicate the experiments of the paper titled [TODO: Add Paper Title]. If you find the code useful or if you use it your research, please cite: [TODO: add publication here]. The data for reproduction is available at [TODO: ADD data link].

Initial implementations by Michelle Lienhart (FCNN) and Yuliia Oksymets (FRCNN). Extended and edited by Manuel Milling.

We provide baseline experiments for two neural network architectures here: A custom fully convolutional neural network (using PyTorch) and an implementation of a Faster RCNN, based on detectron2.

• src/ contains all necessary code to reproduce experiments
• gpu_scripts/ contains scripts to run on a slurm-based GPU-cluster with virtual environments
• results/ contains results from the experiments performed for the paper
• data/ is folder designated for the data to be unpacked to
• logs/ is the default folder for log files

Data Set Preparation

• Download the data at [TODO add link], unzip in the data subdirectory, i.e., an example for an existing image location should be data/sequence_independent/data_and_cell_center_annotations/train/20190420/Images_Part_1_Adhesion/Chip03_1-000.tif

FCNN (Fully Convolutional Neural Network)

Please note that paths and model names need to be adjusted appropriately.

• In order to train the FCNN model, run the training script with custom parameters, e.g. python src/train_FCNN.py --data-root data/sequence_independent/data_and_cell_center_annotations/ --log-root log/log_FCNN/ --distance-threshold 16 --peak-radius 16 --epochs 100 --lr-scheduler-steps 40

• For evaluation, use the according script src/compute_FCNN_scores.py, giving the path to the trained model. The model with best performance on the validation set should be available in the directory specified with the --log-root parameter during training. The results for varying distance thresholds will be stored in Run for instance: python src/compute_FCNN_scores.py -i data/sequence_independent/data_and_cell_center_annotations/ -m log/log_FCNN/model_2023-01-25_17-06-40_nn_spec1_Gaussian=0.625_sstep_size=40_num_epochs=100_size=444_lr=0.002nopost-transformbest_validation_loss.pth --distance-threshold 2 4 8 16 32 --peak-radius 8 --objectness-score 60 -p test --result_dir results/fcnn/

• Plot the original images together with predicted and ground truth cell center predictions with src/plot_FCNN_images.py, e.g., python src/plot_FCNN_images.py -i data/sequence_independent/data_and_cell_center_annotations/test/ -m log/log_FCNN/model_2023-01-25_17-06-40_nn_spec1_Gaussian=0.625_sstep_size=40_num_epochs=100_size=444_lr=0.002nopost-transformbest_validation_loss.pth -o fcnn/images/test/

FRCNN (Faster Region-based Convolutional Neural Network)

• First, prepare the data for the bounding box-based object detection task by running python src/bounding_box_preparation.py. The parameter --rescale converts all images to a common size of 2000x2000 pixels: python src/bounding_box_preparation.py --data-root data/sequence_independent/data_and_cell_center_annotations/ --target-path data/sequence_independent/bounding_box_augmented/ --rescale

• Optionally, run the offline data augmentation script. python src/bounding_box_data_augmentation.py --data-root data/sequence_independent/bounding_box_augmented/

• Train the FRCNN with src/train_FRCNN.py. Set parameters for the index of the current experiment learning rate, number of iterations, etc. To consider augmented samples, use arguments --rotations --cutmix --mixup and/or --cutout, which have to be produced in the previous step. python src/train_FRCNN.py --data-root data/sequence_independent/bounding_box_processed/ --model R101-FPN --train --test --num_exp 11 --eval_period 500 --max_iter 35000 --lr 0.001 --output-dir logs/log_FRCNN/

• For evaluation, first run the script for inference src/inference_FRCNN.py and the src/compute_FCNN_scores.py. The final model of the previous training is be stored in --output-dir and will be used. The results for varying distance thresholds will be stored in --result_dir. Run for instance: python src/inference_FRCNN.py --fimgs data/sequence_independent/bounding_box_processed/dt_test.txt --model R101-FPN --num_exp 1 --output-dir logs/log_FRCNN/ python src/compute_FRCNN_scores.py --labels data/sequence_independent/bounding_box_processed/val_cell_cores.json --predictions logs/log_FRCNN/_1_dt_test_predictions.json --distance-thresholds 2 4 8 16 32 --objectness-score 0.5 --result_dir results/frcnn/

• Plot the original images together with predicted and ground truth bounding boxes with src/plot_FRCNN_images.py, e.g., python src/plot_FRCNN_images.py --label-dir data/sequence_independent/bounding_box_processed/ --predictions logs/log_FRCNN/_1_dt_test_predictions.json --plot_dir results/frcnn/images/test/ --input-img-dir data/sequence_independent/bounding_box_processed/test/ --partition test

• Reproduce distance-dependent plot from the paper using the plot_results.py: python src/plot_results.py --result_dir results/

See also: Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019.