File size: 3,324 Bytes
ebe155d 76d2599 a2a43cd c72844f 76d2599 c72844f 2898778 c72844f 76d2599 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 |
---
'[object Object]': null
license: apache-2.0
language:
- en
tags:
- cell segmentation
- stardist
- hover-net
metrics:
- f1-score
pipeline_tag: image-segmentation
library_name: transformers
---
# Model Card for cell-seg-sribd
<!-- Provide a quick summary of what the model is/does. -->
This repository provides the solution of team Sribd-med for NeurIPS-CellSeg Challenge. The details of our method are described in our paper [Multi-stream Cell Segmentation with Low-level Cues for Multi-modality Images]. Some parts of the codes are from the baseline codes of the NeurIPS-CellSeg-Baseline repository,
You can reproduce our method as follows step by step:
### How to Get Started with the Model
Install requirements by python -m pip install -r requirements.txt
## Training Details
### Training Data
The competition training and tuning data can be downloaded from https://neurips22-cellseg.grand-challenge.org/dataset/ Besides, you can download three publiced data from the following link: Cellpose: https://www.cellpose.org/dataset Omnipose: http://www.cellpose.org/dataset_omnipose Sartorius: https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview
## Environments and Requirements:
Install requirements by
```shell
python -m pip install -r requirements.txt
```
## Dataset
The competition training and tuning data can be downloaded from https://neurips22-cellseg.grand-challenge.org/dataset/
Besides, you can download three publiced data from the following link:
Cellpose: https://www.cellpose.org/dataset
Omnipose: http://www.cellpose.org/dataset_omnipose
Sartorius: https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview
## Automatic cell classification
You can classify the cells into four classes in this step.
Put all the images (competition + Cellpose + Omnipose + Sartorius) in one folder (data/allimages).
Run classification code:
```shell
python classification/unsup_classification.py
```
The results can be stored in data/classification_results/
## CNN-base classification model training
Using the classified images in data/classification_results/. A resnet18 is trained:
```shell
python classification/train_classification.py
```
## Segmentation Training
Pre-training convnext-stardist using all the images (data/allimages).
```shell
python train_convnext_stardist.py
```
For class 0,2,3 finetune on the classified data (Take class1 as a example):
```shell
python finetune_convnext_stardist.py model_dir=(The pretrained convnext-stardist model) data_dir='data/classification_results/class1'
```
For class 1 train the convnext-hover from scratch using classified class 3 data.
```shell
python train_convnext_hover.py data_dir='data/classification_results/class3'
```
Finally, four segmentation models will be trained.
## Trained models
The models are in models/.
## Inference
The inference process includes classification and segmentation.
```shell
python predict.py -i input_path -o output_path --model_path './models'
```
## Evaluation
Calculate the F-score for evaluation:
```shell
python compute_metric.py --gt_path path_to_labels --seg_path output_path
```
## Results
The tuning set F1 score of our method is 0.8795. The rank running time of our method on all the 101 cases in the tuning set is zero in our local
workstation. |