C2FNet

This repository is code for papaer "Weakly-Supervised Nucleus Segmentation Based on Point Annotations: A Coarse-to-Fine Self-Stimulated Learning Strategy", which has been accepted by MICCAI 2020.

Dependencies

Tensorflow 1.15.0

Keras 2.3.0

Usage

Train

split data to train and val set, each set has img and mask folders.

example: 
./data/monuseg/train_val/
|-- train
|   |-- img
|   |-- mask
|-- val
|   |-- img
|   |-- mask

run train_one_fold.py to train segmentation model.
- set in_dataset_fold=train_val, in_dataset_name=monuseg, save_checkpoint_path=./checkpoints/monuseg_ln
- set train_full_mask_flag=True to train fully supervised model
- set itr_sum=4, which indicate that the model will train in 4 iteration, 0,1,2 are in first stage, 3 is in the second stage

Test

run test_edge_point.py to predict result with trained model.
- set fold=train_val,
- set model_name=LinkNet.nuclei.train_val.512_loss_0.01_0.01_0.01_0.01_1.0_train_val_r3_resume_point_edge_fake_sobel.last.h5
- set val_dir=data/monuseg/train_val/val/img/
- set save_dir=data/monuseg/train_val/val/result_r3/
- model_name and save_dir are corresponding, including r0, r1, r2, r3

Evaluation Metrics

cd experiments, and run compute_metrics.py to compute evaluation metrics.
- set base_dir=../data/monuseg/train_val/val/
- set pred_sub_dirs=['result_r0', 'result_r1', 'result_r2', 'result_r3']

Citation

If you find this code helpful, please cite our work:

Tian K, Zhang J, Shen H, et al. Weakly-Supervised Nucleus Segmentation Based on Point Annotations: A Coarse-to-Fine Self-Stimulated Learning Strategy[C] //International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2020: 299-308.