Introduction

Medical Image Segmentation Engine.

Installation

This codebase is only tested on Linux (Ubuntu).

git clone https://github.com/qinliuliuqin/Medical-Segmentation3d-Toolkit.git
cd Medical-Segmentation3d-Toolkit
pip install -e .

Do not forget the last '.' that indicates the current folder.

Usage

1. Inference: Print the help info.

   Input:

   $ seg_infer --help
   

   Output:

   usage: seg_infer [-h] [-i INPUT] [-m MODEL] [-o OUTPUT] [-n SEG_NAME]
                 [-g GPU_ID] [--save_image] [--save_prob]
   
   Inference engine for 3d medical image segmentation It supports multiple kinds
   of input: 1. Single image 2. A text file containing paths of all testing
   images 3. A folder containing all testing images
   
   optional arguments:
     -h, --help            show this help message and exit
     -i INPUT, --input INPUT
                           input folder/single image/txt file
     -m MODEL, --model MODEL
                           model root folder
     -o OUTPUT, --output OUTPUT
                           output folder for segmentation
     -n SEG_NAME, --seg_name SEG_NAME
                           the name of the segmentation result to be saved
     -g GPU_ID, --gpu_id GPU_ID
                           the gpu id to run model, set to -1 if using cpu only.
     --save_image          whether to save original image
     --save_prob           whether to save all prob maps
   
   

   The following is an example that shows how to get started: If you have an image image.mha, a segmentation model model, and a gpu with device id 0, you can run the following command for inference:

   $ seg_infer -i ./image.mha -m ./model -o ./result_folder -g 0 
   

   For inference using CPU, just set gpu id to -1. This package supports different input image types, including .nii.gz, .mhd, .mha, and DCIOM folder. For the DICOM folder, all .dcm files should in this folder.

   WARNING: If you run $ seg_infer with no parameters, the program will crash because it will look for the default paths which are set according to my environment.

   Specifically, if you are working on dental project and if want to segment bony structures from CBCT/CT images, you can download pretrained segmentation models on github (the latest model is model_0429_2020).

   $ git clone https://github.com/qinliuliuqin/Model-Zoo.git  
   

2. Training:

   Firt, configure the training settings in config/train_config.py and then create a training file titled train.txt. An example train.txt should be the following format:

   2
   /home/qinliu/train_data/image_1.mha
   /home/qinliu/train_data/label_1.mha
   /home/qinliu/train_data/image_2.mha
   /home/qinliu/train_data/label_2.mha   
   

   The first number 2 denotes the pair of traning images and their corresponding labels. Use absolute path. You need to set a lot parameters in the configuration file config/train_config.py. Please read the comment in the configuration file carefully.

   Then, run the following code to start training:

   seg_train -i ./config/train_config.py

Requirements

Pytorch=1.3.0 Numpy=1.17.2 SimpleITK=1.2.3