Emphysema Subtyping on Chest CT using Deep Neural Networks

News

• The source code (v1.0) is now available.

Table Of Contents

• Introduction
• Usage
• Main Results

Introduction

The proposed model can automatically identify severity-based emphysema subtypes according to Fleischner visual scoring system by analyzing a given CT scan. The proposed model outperformed the existing method on the presented dataset with improved interpretability.

Usage

• Use train.py for training. The training, testing and prediction scripts were all implemented using pytorch, and pytorch-lightning library.
• Use processor.py or run.sh for inference. both processor.py and run.sh require you define the input data (ct image and its lobe segmentation) using --scan_path and --lobe_path, and the output path using --output_path arguments.
• The code supports inference and training using multiple GPUs. Please use --ngpus and --workers to specify the number of GPUs and the number of workers for the executation. Check line 60 and 70 for details of possible input arguments.
• Please check \install_files\requirements.in for 3rd-party libraries to be installed to run the scripts. Run pip install -r install_files/requirements.in to install dependencies. The code has been tested with python 3.8 version. If you want to install torch with GPU support, please use --extra-index-url=https://download.pytorch.org/whl/cu113 (chose the cuda version you have already installed, e.g., 11.3 in this example).
• We provide the classification and regression training strategies. Please switch to med3d in --model_arch cli argument.
• The class and regression activation maps were generated during training or testing.
• For the Grand-challenge algorithm, we use the prediction mode in pytorch-lightning for outputs.

Main Results

Tab 1. Centrilobular and ParaseptalEmphysema Severity Scores Classification Accuracy (ACC(%)) and F-measurement, in comparison with the Fleischner algorithm.

Method	Subtype	ACC (%)	F1-score	Linear Weighted Kappa(95% CI)
The Fleischner algorithm	CLE	45	-	60
Ours (classification)	CLE	52.23	51.00	64.29 (63.16-65.42)
Ours (classification)	PSE	59.12	57.12	42.03 (40.21-43.85)
Ours (regression)	CLE	51.32	49.61	64.24 (63.14-65.35)
Ours (regression)	PSE	64.62	60.74	52.06 (50.40-53.73)

Qualitative Results

The first row showcases the dense regression activation maps (dRAM) for centrilobular emphysema, and the second row illustrates the dRAM for paraseptal emphysema.

MIT

License

MIT