jaeho3690 / lidc-idri-segmentation

This is the segmentation process of the LIDC-IDRI dataset. Checkout my preprocessing repository to use this repository

LIDC Segmentation of Lung Nodule with U-Net, U-Net++

This repository is the second stage for Lung Cancer project. Please check out my first repository LIDC-IDRI-Preprocessing Explanation for my first repository is on Medium as well! The input for this repository requires the output format from the first stage. This repository would train a segmentation model(U-Net, U-Net++) for Lung Nodules. The whole script is implemented in Pytorch Framework. The model script for U-Net++ and some of the script format is sourced from here

Requirements

• pytorch 1.4
• GPU is needed

1. Check out my LIDC-IDRI-Preprocessing repository

This repository goes through the preprocessing steps of the LIDC-IDRI data. Running the script will return .npy images for each lung cancer slice and mask slice. Also, a meta.csv, clean_meta.csv file will be made after running the jupyter file.

2. Fix directory settings

All the scripts were written when I was not so familiar with directory settings. I mostly used absolute directory. Please change each directory setting to fit yours. I apologize for the inconvenience.

Installation

1. Create a virtual environment

conda create -n=<env_name> python=3.6 
conda activate <env_name>

2. Install pip packages

pip install -r requirements.txt

File Structure

+-- Unet
|    # This folder contains the model code for U-Net
+-- UnetNested
|    # This folder contains the model code for U-Net++
+-- figures
|    # This folder saves figure images
+-- meta_csv
|    # This folder contains information of each images in a csv format. 
|    # The cs
+-- notebook
|    # This folder contains jupyter notebook files for some visuialization
+-- dataset.py
|    # Dataset class for Pytorch, Accepts .npy file format
+-- losses.py
|    # Loss function. Here I use the BCE dice loss. Sourced from 
+-- metrics.py
     # Metric function. It is interesting to note that somehow the dice coefficient doesn't increase as fast as IOU in the early stages of training.
+-- train.py
|    # Training of Segmentation model. Adjust hyperparameters
+-- utils.py
|    # Utility file
+-- validate.py
|    # For validation of the model

Training

1. Train the model. There will be total of 4 cases. UNET_base, UNET_with_augmentation, NestedUNET_base, NestedUNET_with_augmentation

# Training U-net
python train.py --name UNET --augmentation True
# Training U-Net++
python train.py --name NestedUNET --augmentation True 

2. Validate the model

# if you want to get the augmented version
python validate.py --name UNET --augmentaton True

Some resources that were useful

Evaluatuion Metric https://www.sciencedirect.com/science/article/pii/S1361841510000587?via%3Dihub

lidc-idri-segmentation's Issues

About num_class

Hello, I am a beginner, I would like to ask why num_class=1, not 2, in your code, isn't the image segmentation task usually set to 2

About the segmentation performance

Thanks for your work! I would like to ask what is your final segmentation performance? The dice I get on the validation set is only around 0.5. Looking forward to your reply! Thank you! : )

"data_split" is missing

There is no 'data_split' in meta.csb file.

CUDA Memory Error

Hi Jaeho,
It's a very nice repository for segmentation model. I'm using your repository for segmenting the data of LIDC-IDRI but I'm facing the following error while running your codes as per your directions:

I'm using NVIDIA Geforce 940MX GPU with cuda, pytorch installed in virtual env.
Can you please help, what could be the probable solution.
Thanks.

crop_nodule function

I am unable to find the function crop_nodule in View_output.ipynb.
The function crop_nodule is called in function crop_patch.

Dice score coefficient

Hi Jaeho,

I was wondering what results you were able to get with your segmentation model? Specifically UNET (with and without augmentation).
Please let me know!