This repository contains code for combining 3D convolutional neural networks (CNNs) and support vector machines (SVMs) to predict the longitudinal progression of lung decline.

• data.ipynb contains statistics on clinical biomarker data and CT scan data. It also includes code for segmenting lung volumes from volumetric CT scans and adjusting the resolution of CT scans.
• result.ipynb contains codes for generating all results, including SVM and CNN results with varying resolutions, losses, and hyperparameters, as well as hybrid model and importance map results.
• figure.ipynb contains the code necessary to reproduce all figures presented in the paper.

All Python files are located inside the src directory.

• train_cnn.py is the primary code for training a 3D CNN to predict lung function decline.
• segmentation.py is used to extract only the lung volumes from volumetric CT scans.
• saliency.py calculates the saliency map, which is the Integrated Gradients of the 3D CNN prediction.
• data_utils.py contains functions to process and load data.
• result_utils.py includes functions for calculating the results of each model.
• hybrid_utils.py includes functions for hybrid modeling, such as cascade, and-, or-, and avg-voting.
• sal_utils.py contains a function to visualize 3D saliency maps.
• aux_utils.py contains some helping functions.

To train a 3D CNN, you can

./train_cnn_gpu.sh

The dataset utilized in this study can be accessed by submitting a formal request to the SARP Data Coordinating Center.