Nuclei-segmentation

This repository gathers the code for nuclei-segmentation from the in-class CodaLab competition.

We use Detectron2, a Python API provided by Facebook research for Mask R-CNN, based on the PyTorch framework, to train our model. In this competition, we use Mask R-CNN on ResNet-101 and ResNeXt-101(32x8d) backbone these two models and analysis the results.

Reproducing Submission

We need to do some pre-preparation for training and testing on our custom dataset.

To reproduce my submission without retrainig, do the following steps:

1. Requirement
2. Repository Structure
3. Dataset setting
4. Inference

Hardware

Ubuntu 20.04.3 LTS

Intel® Core™ i7-9700KF CPU @ 3.60GHz × 8

GeForce RTX 3090 32G

Requirement

All requirements should be:

$ virtualenv detectron2 --python=3.8
$ source ./detectron2/bin/activate
$ cd Nuclei-segmentation
$ pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 -f \
  https://download.pytorch.org/whl/torch_stable.html
$ python -m pip install detectron2 -f \
  https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.9/index.html
$ pip install opencv-python
$ pip install scikit-image

Official images can be downloaded from CodaLab competition

Repository Structure

The repository structure is:

Nuclei-segmentation(root)
  +-- configs                    # all configs(hyperparameters) used in the program 
  |   +-- COCO-InstanceSegmentation
      |   +-- mask_rcnn_R_101_FPN_3x.yaml
      |   +-- mask_rcnn_X_101_32x8d_FPN_3x.yaml
  |   +-- Base-RCNN-FPN.yaml
  +-- dataset                    # all training and testing files
  |   +-- coco
      |   +-- annotations        # json files
          |   +-- test_img_ids.json   
          |   +-- test.json   
          |   +-- train.json   
      |   +-- test               # testing set .png  
          |   +-- TCGA-50-5931-01Z-00-DX1.png  
          |   +-- TCGA-A7-A13E-01Z-00-DX1.png 
          |   +-- ......
      |   +-- train              # origin training data (each image have one folder)
      |   +-- trainval           # training set .png
          |   +-- TCGA-18-5592-01Z-00-DX1.png   
          |   +-- TCGA-21-5784-01Z-00-DX1.png 
          |   +-- ...... 
      |   +-- parse_test.py      # parse testing json file
      |   +-- parse_train.py     # parse training json file
  +-- input_img                  # save training images with segmentation there
  +-- output_img                 # save testing images with segmentation there
  +-- output                     # save training model and matrics.
  +-- visual_test.py             # output testing images with segmentation
  +-- visual_train.py            # output training images with segmentation
  +-- inference.py               # model prediction and reproduce my submission file
  +-- train.py                   # for training model

Dataset setting

You can use parse_train.py to parse training images to .json file and also use parse_test.py to parse testing images to .json file.

$ python dataset/coco/parse_test.py
$ python dataset/coco/parse_train.py

Training

To train the model, run this command:

$ python train.py

Trained model will be saved as output/R101/{1, 2, 3,...}/model_final.pth

All mAP of experiments will be written in Results.

Inference

Please download this model if you want to reproduce my submission file, put it in root/model_final.pth and run the following code.

To reproduce my submission file or test the model you trained, run:

$ python inference.py

Prediction file will be saved as root/answer.json

If you use different hardware, the inference result may be a little different.

Results

Our model achieves the following performance on :

ResNet-101 Result:

No.	1	2	3	4	5	6	7	8	9	10	11
model	R101	R101	R101	R101	R101	R101	R101	R101	R101	R101	R101
mAP	0.223274	0.219256	0.208042	0.221926	0.226062	0.211141	0.242515	0.24355	0.232426	0.233071	0.242185
Batch size	128	512	128	128	128	128	128	128	128	128	128
Anchor size	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	8, 16, 32, 64, 128	16, 32, 64, 128, 256
Iteration	300000	300000	300000	500000	50000	200000	6000	5000	4000	5000	5000
Image size	800	800	1200	1200	1200	1200	1200	1200	1200	1200	1200
Note(change)		batch size	image size	iteration	iteration	iteration	iteration	iteration	iteration	anchor size	anchor size

ResNeXt-101(32x8d) Result:

No.	12	13	14	15
model	X101	X101	X101	X101
mAP	0.228375	0.225095	0.243045	0.231643
Batch size	128	128	128	128
Anchor size	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512	32, 64, 128, 256, 512
Iteration	50000	100000	5000	3000
Image size	1000	1000	1000	1000
Note		iteration	iteration	iteration

Reference

[1] Detectron2

[2] pycocotools

[3] RLE encode