Reproducible Machine Learning Methods for Lung Cancer Detection Using Computed Tomography Images: Algorithm...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Computational Pathology

Background:

• Chest computed tomography (CT) is vital for lung cancer detection, yet automated evaluation methods face reproducibility challenges due to diverse software dependencies.
• Lack of standardized comparison and reproduction hinders progress in automated lung cancer detection using CT scans.

Purpose of the Study:

• To generate reproducible machine learning modules for lung cancer detection.
• To compare the approaches and performance of award-winning algorithms from the Kaggle Data Science Bowl challenge.

Main Methods:

• Source codes from award-winning Kaggle Data Science Bowl solutions for lung cancer detection were obtained.
• Algorithm performance was evaluated using the log-loss function and Spearman correlation coefficient.
• Docker containers were generated for top solutions to ensure reproducibility.

Main Results:

• Commonly used methods included U-Net, VGGNet, and residual networks for segmentation, with transfer learning prevalent in classification.
• Significant performance variations were observed across different test sets (Spearman correlation coefficient = .39 among top 10 teams).
• Reproducible Docker images were successfully generated for the top lung cancer detection algorithms.

Conclusions:

• A comparative analysis of award-winning lung cancer detection algorithms was performed, yielding reproducible Docker images.
• Convolutional neural networks demonstrated good accuracy, but enhancing model generalizability is crucial for clinical application.