NaviAirway

NaviAirway: a Bronchiole-sensitive Deep Learning-based Airway Segmentation Pipeline, Preliminary version presented at RSNA2021.

Airway segmentation is essential for chest CT image analysis. However, it remains a challenging task because of the intrinsic complex tree-like structure and imbalanced sizes of airway branches. Current deep learning-based methods focus on model structure design while the potential of training strategy and loss function have not been fully explored. Therefore, we present a simple yet effective airway segmentation pipeline, denoted NaviAirway, which finds finer bronchioles with a bronchiole-sensitive loss function and a human-vision-inspired iterative training strategy. Experimental results show that NaviAirway outperforms existing methods, particularly in identification of higher generation bronchioles and robustness to new CT scans. Besides, NaviAirway is general. It can be combined with different backbone models and significantly improve their performance. Moreover, we propose two new metrics (Branch Detected and Tree-length Detected) for a more comprehensive and fairer evaluation of deep learning-based airway segmentation approaches. NaviAirway can generate airway roadmap for Navigation Bronchoscopy and can also be applied to other scenarios when segmenting fine and long tubular structures in biomedical images.

Pipeline

Demonstration (code, data, and instruction)

OneDrive password: 2333

Dependencies

Check the required python packages in requirements.txt.

Datasets

EXACT09

The file structure should be like this

/data/Airway/EXACT09
    /Training
        /CASE01
            /1093782
            /1093783
            ...
        /CASE02
        ...
    /Testing
        /CASE21
        ...

LIDC-IDRI

The file structure should be like this

/data/Airway/LIDC-IDRI
    /LIDC-IDRI-0001
        /1.3.6.1.4.1.14519.5.2.1.6279.6001.298806137288633453246975630178
            /1.3.6.1.4.1.14519.5.2.1.6279.6001.179049373636438705059720603192
                /1-001.dcm
                /1-002.dcm
                ...
    /LIDC-IDRI-0002
    ...

Run the code

[Note] Folder dataset_info contains data info used for training scripts. You can read the sample files in that folder to learn the format.
[Dataset preparation]
- Download the two datasets: EXACT09 and LIDC-IDRI.
- Run dataset_preprocess_EXACT09.ipynb and dataset_preprocess_LIDC-IDRI.ipynb to preprocess the images.
- Run Pre_crop_images.ipynb to pre-crop the images to be samll cubes.
- Run Get_dataset_info.ipynb to generate dataset info which are pkl files (our iterative training strategy (training with focus on airways of low and high generations iteratively) is achieved by it).
[Training] Run train.py or train_semi_supervised_learning.py to start training. You can change the hyperparameters. Model parameters are saved in checkpoint.
[Inference] Run NaviAirway_pipeline.ipynb.

Results

See the 3D models in results.

Contact

If you have any questions, please contact wangad@connect.hku.hk.

qgking / NaviAirway