AeroPath: An airway segmentation benchmark dataset with challenging pathology and baseline method.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Accurate airway segmentation in CT scans is vital for diagnosing and treating lung diseases, including cancer.
• Existing datasets like ATM'22 have advanced airway segmentation but lack sufficient examples of severe airway pathologies.
• Deep learning models show promise for automated airway segmentation, but require diverse datasets for robust performance.

Purpose of the Study:

• Introduce the AeroPath dataset, a new public benchmark with CT scans and airway annotations from patients with diverse and severe lung pathologies.
• Present a novel multiscale fusion deep learning architecture for robust and accurate automatic airway segmentation.
• Evaluate the proposed model's performance against existing methods and provide an accessible web application for testing.

Main Methods:

• Developed the AeroPath dataset comprising 27 CT scans with detailed trachea and bronchi annotations for pathologies like emphysema and tumors.
• Designed and implemented a multiscale fusion network for automatic airway segmentation.
• Trained models on the ATM'22 dataset and tested on the AeroPath dataset, benchmarking against competitive open-source methods using ATM'22 challenge metrics.
• Created an open web application for easy model testing on new data.

Main Results:

• The proposed multiscale fusion model achieved topologically correct airway segmentations for all patients in the AeroPath dataset.
• The method demonstrated robustness in handling various airway anomalies down to the fifth generation.
• Performance was benchmarked against leading open-source methods, showing competitive or superior results.
• The AeroPath dataset proved valuable for evaluating methods on challenging pathologies.

Conclusions:

• The AeroPath dataset and the proposed multiscale fusion model significantly advance the field of airway segmentation, particularly for complex lung pathologies.
• The developed method offers a robust solution for airway segmentation, aiding in early diagnosis and intervention planning for pulmonary diseases.
• Openly sharing the AeroPath dataset and web application will foster further research and development in medical image analysis for respiratory diseases.