kxdkxd/laryngocarcinoma_segmentation

laryngocarcinoma_segmentation, with OCRNet on 2D z-axis slices

laryngocarcinoma_segmentation

Laryngocarcinoma Segmentation on CT Nii files, with OCRNet on 2D z-axis slices, implemented and tested on Python3 and Paddle 2.0.2

Note: Paddle framework is much as similar as PyTorch.

Before Run

Basic System Environment

Hardware

For Training

• Tested on Baidu AIStudio, with NVIDIA Tesla V100 - 32GiB VideoMem. While training, around 24GiB VideoMem will be used with batch_size is set to 12. Lower the batch_size can significantly reduce the VideoMem cost, but as an important hyper-parameter, this move can lead to different performance as I do.

• As a result, a GPU with large VideoMem is recommended.

For Testing

• Tested on Baidu AIStudio, with pure CPU, really slow and ended up with Error. However works great with a GPU.

• A Normal GPU can handle testing job.

Software

• Python 3 (Tested on Python 3.7)
• Paddle 2.0.2
• paddleseg (Tested on pip installed paddleseg 2.2.0)
• nibabel (Tested on pip installed nibabel 3.2.1)
• ITK-SNAP (3.6.0) (For 3D nii data visualization)

Tutorial

Install packages:

pip install nibabel paddleseg

Download OCRNet pre-train model:

wget https://bj.bcebos.com/paddleseg/dygraph/cityscapes/ocrnet_hrnetw48_cityscapes_1024x512_160k/model.pdparams -o model_pretained_bcebos.pdparams

Construct Directories:

mkdir dataset
mkdir dataset/srcs
mkdir dataset/labels
mkdir dataset/labels_show
mkdir outputs
mkdir results

Download dataset:

Download from BaiduNetdisk, place TrainSet.zip TrainLabel.zip to dataset directory. Place outputs.zip to the project root directory if you just want to perform test job.

https://pan.baidu.com/s/1NZ0UyAq5c_dno-N1pojSVA  code: kkkk

unzip outputs.zip

cd dataset/ && unzip TrainSet.zip && unzip TrainLabel.zip

Note: The dataset-08 is mal-formed. The label should be {0, 1} while {2} occured, and the volume of labeled data is abnomarlly large. Pay attention to this one.

The given best model can achieve mIoU: 0.8355 Acc: 0.9991 Kappa: 0.8040 ,Class IoU: [0.9991 0.6719], Class Acc: [0.9996 0.7973] on given dataset.

Run Train & Validate

Preprocess dataset:

python preprocess_for_train.py

dataset divided train/val = 9/1. You can Edit this line.

test_data_size = int(len(data_list) * 0.1)

Train

python train.py --config OCRNet_W48.yml --do_eval --save_interval 50 --save_dir outputs

Validate

python val.py --config OCRNet_W48.yml --model_path outputs/best_model/model.pdparams

Run Test

Preprocess dataset:

python preprocess_for_test.py

python predict.py --config OCRNet_W48.yml --model_path output/best_model/model.pdparams --image_path dataset/srcs --save_dir results

The test result is also in 2D z-axis sliced. If you want to reformat them into NII format, try MATLAB dicm2nii. You can access this repo on github.

clc; clear;
files = dir('*.jpg')
file_names = {files.name}
num_of_files=numel(file_names);
for i=1:num_of_files
  nifti_mat(:,:,i)=imread(file_names{i});
end
nii = nii_tool('init', nifti_mat);
nii_tool('save', nii, 'label90.nii'); %%% NUMBER %%%

Appendix

Since I only introduced 2D slice samples with a labeled tumor. The model may predict a tumor at a wrong part of a body, if you input a full-body-part CT Data, for example mistakenly predict a tumor in brain or bowel. These part should be carefully examine after finishing predict job. One method to remove these mistakes is open the origin nii main image in ITK-SNAP and compact these predicted 2D files into nii 3D files with dicm2nii, then import to ITK-SNAP as segmentation. Press Update Button on the left-down corner to observe 3D-Model. Then note the wrong z-axis slice number, then back to jpgs folders, replace the determinded number of slice with another absolute 0 label jpg. Then re-run the dicm2nii again to generate final NII result.