MitoEM2021-Challenge

Here is the official code of the 1st team "VIDAR" on the MitoEM Challenge.

More detial can refer to the paper, Advanced Deep Networks for 3D Mitochondria Instance Segmentation and the report on the MitoEM challenge.

🏆 SOTA for 3D Instance Segmentation on MitoEM. Check out Papers With Code for the 3D Instance Segmentation on MitoEM and more details.

We publish the validation results of our model on the rat and the human at Google Drive.

Dataset

	Validation set	Validation set	Test set
MitoEM-R	400×4096 ×4096	100×4096 ×4096	500×4096 ×4096
MitoEM-H	400×4096 ×4096	100×4096 ×4096	500×4096 ×4096

Framework

Installation

You may refer to the installation.
The additional packages can be installed as:

pip install torchsummary waterz malis

We also give you a docker (Ali Cloud and Dockerhub, cuda9.0_torch1.1) , you can pull it to your server.

	source
Ali Cloud	registry.cn-hangzhou.aliyuncs.com/ustc-team/cuda_9.0_pytorch1.1:v2
Dockerhub	dockerlimx/cuda_9.0_pytorch1.1:v2

Training stage

Take rat for example, in rat/configs/MitoEM/, you need to modify "im_train_rat.json", "mito_train_rat.json" for training sample and "im_val_rat.json" for validation sample and "im_test_human.json", "im_test_rat.json" for test sample. Note the validation GT file is saved as "h5", you can use the code to convert the image slices into an H5 file. Because the challenge calculates the "h5" input for evaluation.

GPU memory friendly

We train the network for a certain interval by default, and use cal_infer to inference and evaluate the predicted results. If you have enough resources, you can use this function during training. Otherwise, please use offline testing.

Training setting

Take rat for example, you can change "rat/configs/MitoEM/MitoEM-R-BC.yaml". Specificly, you may change the INPUT_PATH and INFERENCE_PATH.

SYSTEM:
  NUM_GPUS: 2
  NUM_CPUS: 8
MODEL:
  ARCHITECTURE: 'rsunet'
  FILTERS: [28, 36, 48, 64, 80]
  INPUT_SIZE: [36, 320, 320]
  OUTPUT_SIZE: [36, 320, 320]
  IN_PLANES: 1
  OUT_PLANES: 2
  LOSS_OPTION: [['DiceLoss', 'WeightedBCE'], ['DiceLoss', 'WeightedBCE']]
  LOSS_WEIGHT: [[0, 1], [0, 1]]
#  LOSS_OPTION: [['WeightedBCE'], ['WeightedBCE']]
#  LOSS_WEIGHT: [[1], [1]]
  TARGET_OPT: ['0','4-2-1'] # Multi-task learning: binary mask, instance segmentation
  WEIGHT_OPT: [['1'],['1']]
DATASET:
  IMAGE_NAME: 'configs/MitoEM/im_train_rat.json'
  LABEL_NAME: 'configs/MitoEM/mito_train_rat.json'
  INPUT_PATH: '/braindat/lab/limx/MitoEM2021/CODE/HUMAN/rsunet_retrain_297000_v2/' # work container
  # inference: save model
  INFERENCE_PATH: '/braindat/lab/limx/MitoEM2021/CODE/HUMAN/rsunet_retrain_297000_v2/'
  OUTPUT_PATH: 'outputs/dataset_output'
  PAD_SIZE: [0, 0, 0] # [16, 128, 128]  # Mirror padding of big chunk
  DO_CHUNK_TITLE: 1 # json file reading
  DATA_CHUNK_NUM: [8, 2, 2] # [8, 2, 2] # block number of each axis.
  DATA_CHUNK_ITER: 2500 # sample times of per chunk
  LABEL_EROSION: 1
  USE_LABEL_SMOOTH: False
  LABEL_SMOOTH: 0.1

SOLVER:
  LR_SCHEDULER_NAME: "WarmupMultiStepLR"
  BASE_LR: 1e-04
  ITERATION_STEP: 1 # How many iterations return loss once
  ITERATION_SAVE: 1500  # save model
  ITERATION_TOTAL: 300000 # total iteration
  SAMPLES_PER_BATCH: 2 #
INFERENCE:
  INPUT_SIZE: [32, 256, 256]
  OUTPUT_SIZE: [32, 256, 256]
  IMAGE_NAME: 'configs/MitoEM/im_val_rat.json'
  OUTPUT_PATH: 'outputs/inference_output'
  OUTPUT_NAME: 'result.h5'
  PAD_SIZE:  [16, 128, 128]
  AUG_MODE: 'mean'
  AUG_NUM: 0
  STRIDE: [16, 128, 128] # [16, 128, 128]
  SAMPLES_PER_BATCH: 48

Run the main.py to start and the results will be saved in rat/outputs/inference_output

cd rat &&\
python scripts/main.py --config-file configs/MitoEM/MitoEM-R-BC.yaml

If you have checkpoint file:

cd rat &&\
python scripts/main.py --config-file configs/MitoEM/MitoEM-R-BC.yaml --checkpoint xxx

Validation stage

By default, AP-75 is calculated on the validation set every certain epoch during the training stage. In our experiment, the validation stage needs almost 45 GB of memory, which is mainly due to post-processing for the seed map of size 100 × 4096 × 4096. Therefore, we recommend you use 64GB of memory during the training stage.

Similarly, you can refer to eval_iter.sh to quickly run inference and evaluation programs.

And you can refer rat/connectomics/utils/evaluation/iteration_eval.py to call the inference and evaluation. Please check it and confirm it works.

Segmentation results

	AP-75 (Validation set)
Res-UNet-R+MT	0.917
Res-UNet-H+MT	0.828

We publish the validation results of our model on the rat and the human at Google Drive.

Model Zoo

Download the pretrained models here.

Model	Dataset
checkpoint_085000.pth.tar	Res-UNet-R
checkpoint_297000.pth.tar	Res-UNet-H

Please change ARCHITECTURE in config file.

Acknowledgement

This project is built upon numerous previous projects. Especially, we'd like to thank the contributors of the following github repositories:

pyGreenTea: HHMI Janelia FlyEM Team
DataProvider: Princeton SeungLab
Detectron2: Facebook AI Reserach
pytorch_connectomics: Harvard University

Citation

If you find this work or code is helpful in your research, please cite:

@article{li2021advanced,
title={Advanced Deep Networks for 3D Mitochondria Instance Segmentation},
author={Li, Mingxing and Chen, Chang and Liu, Xiaoyu and Huang, Wei and Zhang, Yueyi and Xiong, Zhiwei},
journal={arXiv preprint arXiv:2104.07961},
year={2021}
}

smittenkittenz / MitoEM2021-Challenge