Lung Tumor Segmentation

A project I'm currently working on just for fun and educational purposes. In this project, I used the lung tumor data from the Medical Decathlon competition

Using this project

In order for this project to log things properly, you need to:

• run wandb login in your terminal and provide your own key

To run this project yourself:

Python 3.7 is required.

1. Clone the project, cd into project directory and run pip install -e .
2. Install the packages from the requirements.txt file in your python env.
3. Download the medical decathlon lung tumor data and extract it to a folder.
4. Run: python preprocessing.py --input_data_dir <extraction_path>/imagesTr --input_labels_dir <extraction_path>/labelsTr --output_dir <path_to_output_dir> (If you want to try this out on your own data - the expected format is Nifty for both the scan and the mask data. The scan and corresponding mask must have the same name and be in different folders)
5. Run: python train.py --preprocessed_input_dir <path_to_output_dir> (This is the output directory you provided in the previous step)
6. Inference : run python inference.py --path_to_ckpt <path_to_ckpt> --path_to_ct_scan <path_to_nifty_ct_scan> --path_to_result_dir <output_dir>.

My SegNet checkpoint can be downloaded from this link and used with the inference script.

Results

So far, using the architecture proposed in the SegNet paper, I reached nice results (0.88 dice score, 0.75 IoU on the validation set)

Example result:

Here is the prediction for an entire ct scan

predicted.mp4

As we can see, it's not accurate but still a nice result 😄

It is still a work in progress