naivete5656 / mdpafof

Mitosis Detection from Partial Annotation by Dataset Generation via Frame-Order Flipping, in MICCAI 2023

Mitosis Detection from Partial Annotation by Dataset Generation via Frame-Order Flipping (MICCAI 2023)

Introduction

This is an official pytorch implementation of Mitosis Detection from Partial Annotation by Dataset Generation via Frame-Order Flipping (MICCAI 2023).

Main Results

Results on 4 datasets

Method	L	HeLa	ES	ES-D	FIb	Ave.
Ours	1 annot.	0.593	0.740	0.439	0.440	0.553
Ours	5 annot.	0.795	0.843	0.628	0.451	0.610
Ours	F	0.841	0.870	0.648	0.541	0.725
Supervised	F	0.868	0.812	0.630	0.304	0.653

Environments

• Ubuntu 20.04.6 LTS
• python >= 3.6
• NVIDIA RTX A6000

Installation

Conda user

conda env create -f=requirement.yml
conda activate pytorch

Docker user

bash ./docker/build.sh
bash ./docker/run.sh

Data preparation

Celltracking challenge

Please download training data and test data of "HeLa cells stably expressing H2b-GFP". Extract them under {ROOT_PATH}/original_data. Please rename test data file as Fluo-N2Dl-HeLa_test. Please run preprocessing code. The code normalize image based on max and mix value of whole seqence.

python ./pre_processing/ctc_preprocess.py --root_paht {ROOT_PATH}

${ROOT_PATH}
|- data
    |-- ctc_preprocessed
        |-- Fluo-N2DL-HeLa
        |   |-- 01
        |   |   |- imgs
        |   |   |  |- t000.tif
        |   |   |      :
        |   |   |  L  t091.tif
        |   |   L 01_mit.txt
        |   |
        |   L   02
        |       |- imgs
        |       |  |- t000.tif
        |       |      :
        |       |  L  t091.tif
        |       L 02_mit.txt
        L   Fluo-N2DL-HeLa_test
            |-- 01
            L   02
        

For the few-shot experiments and the missing experiments, please use sampling code in pre_processing folder.

Training and Testing

Training and test script for 4-fold cross validation on 5 seed conditions (Our experiments)

bash ./scripts/croos_train.sh

Train on one condition.

python ./scripts/cross_train.py --img_dir ./datas/ctc_preprocessed/Fluo-N2DL-HeLa\
        --cv_num 0 --seed 0 --shot 5\
        --save_path ./outputs/${cell_type}/${method}/shot${shot}/seed${seed}/${cv_num}\
        --weight_path ./weights/${cell_type}/${method}/shot${shot}/seed${seed}/${cv_num}/best.pth\
        --dataloader CVCPLoaderBrend2

Citation

If you use our code or models in your research, please cite with:

@inproceedings{Nishimura2023Mitosis,
  title={Mitosis Detection from Partial Annotation by Dataset Generation via Frame-Order Flipping},
  author={Nishimura, Kazuya and Katanaya, Ami and Chunma, Shinichiro and Bise, Ryoma},
  booktitle={MICCAI},
  year={2023}
}