⚠️ Warning: Performance of the system is still incomplete and far away from practical use.

• Application Demo
• Introduction
• Models
• Built With
• Getting Started
• Future Work
• Acknowledgments

This is a pulmonary nodules (lung nodules) computer-aided diagnosis (CADx) demo application, which cound directly raise a diagnostic suggestion based on Lung-RADS automatically from computed tomography (CT) scan.

The application has three different view:

• Main Window
• Screening Window
• Preferences

• Have a quick view of patient list and corresponding scans & nodules information.
• Field introduction
  • Upper block: Patient list
    • Updated: Indicate if all the scans of this patient have been detected
    • Patient Name/Date of Birth/Sex : Information of patient
    • Score/Management: Output based on Lung-RADS
  • Lower block: Scan & Nodule list
    • Updated: Indicate if the scan has been detected
    • Scan date/File Directory: Information of scan
    • Type: Texture type of corresponding nodule, has five categories (Non-Solid, Non/Part, Part-Solid, Part/Solid, Solid)
    • Calcification/Spiculation/Perifissural/Endobronchial: 0 indicate No and 1 indicate Yes
    • Score: Score of single nodule based on Lung-RADS
• Button introduction
  • Load Scan: Load CT scan
  • Detect: Run automatic diagnosis on seleteced scan, it may take minutes
  • Save Changes: Since the information of nodules could be edited by user, save those changes to project file (.json)
  • Export Report: Export field information in csv format
  • Display: Open screening window for selected scan

• Showing selected scan and it's corresponding detected nodules
• Since model isn't perfect, user can edit nodules information here

• Several simple settings of application
• LIDC-IDRI dataset label texture in 5 categories (1~5), with 2 & 4 is undetermined. Check the box will classify 2 with 1, and 4 with 5
• Threshold indicate thresholding the probability at detection stage

There are two models in models

• nodulenet_model.ckpt
  • Perform nodule detection & segmentation
  • Follow the training procedure describe at NoduleNet repo
  • Since I modified some data_parallel code in net to make model run on CPU, replace the net directory for your own use may be needed
• classification_model.pkl
  • Perform nodule texture classification (only for Non-Solid, Part-Solid and Solid classification)
  • Train with LIDC-IDRI dataset (LUNA16 1186 nodules selected version)
  • Using Radiomics methodology, with Principal Component Analysis (PCA) and Support Vector Regression (SVR)

Modules besides python built-ins

• GUI - PyQt5 / QDarkStyle / PyQtGraph
• Basic - NumPy / SimpleITK / OpenCV / Scikit-Image / SciPy
• Model Relatives - PyTorch / Scikit-Learn / PyRadiomics
• Others - Faker / dateutil

• Ensure you could run a NoduleNet model on your device
• Replace net directory or config.py or nodulenet_model.ckpt may be needed

• Clone the repo

git clone https://github.com/wenyalintw/Nodule_CADx.git

• TODO Install required modules listed in requirements.txt

pip install -r /path/to/requirements.txt

• Ready to go

cd src
python main.py

• Lung ROI segmentation
• Size based on Lung-RADS criteria
• Types classification beside texture

• NoduleNet repo
• Reference of screening window apperance design: ITK-SNAP
• trashcan.png licensed under "CC BY-ND 3.0" downloaded from ICONSDB