# Automatic Airway Segmentation in chest CT using Convolutional Neural Networks

**Authors:** A. Garcia-Uceda Juarez, H. A. W. M. Tiddens, M. de Bruijne  
**arXiv:** [1808.04576](https://arxiv.org/abs/1808.04576)  
**Date:** 2018-08-14  
**Summary page:** https://www.summarizepaper.com/en/paper/1808.04576v1/

## Key points

- Study focuses on segmentation of airway tree from chest CT images
- Accurate airway segmentation crucial for assessing airway diseases like bronchiectasis and COPD
- Complex nature of airway structures makes precise segmentations challenging
- Researchers propose a simple yet robust approach using 3D Unet architecture
- Deep CNNs, specifically Unet architecture, proven effective for biomedical image segmentation
- Method trained on 12 CT scans and tested on 6 CT scans
- Impact of different loss functions and data augmentation techniques evaluated
- Average dice coefficient of 0.8 between ground-truth and automated segmentations reported
- Promising solution for accurate airway segmentation in chest CT images using deep learning techniques and adapted Unet architecture
- Contributes to advancing quantitative assessment methods for airway diseases with potential implications for diagnosis and treatment planning in clinical settings.

## Summary

This study focuses on the segmentation of the airway tree from chest computed tomography (CT) images, which is crucial for accurately assessing airway diseases such as bronchiectasis and chronic obstructive pulmonary disease (COPD). Obtaining precise segmentations from CT scans has proven challenging due to the complex nature of airway structures. To address this issue, the researchers propose a simple yet robust approach based on a 3D Unet architecture. Deep convolutional neural networks (CNNs), particularly the Unet architecture designed for biomedical images, have emerged as state-of-the-art methods for various segmentation tasks. However, applying these techniques to airway segmentation remains difficult. The method developed in this work is trained on a dataset consisting of 12 CT scans and tested on another 6 CT scans. The researchers evaluate the impact of different loss functions and data augmentation techniques on the performance of their approach. They report an average dice coefficient of 0.8 between the ground-truth segmentations and their automated segmentations. Overall, this study presents a promising solution for accurate airway segmentation in chest CT images by leveraging deep learning techniques and specifically adapting the Unet architecture to handle the complexities of airway structures. This research contributes to advancing quantitative assessment methods for airway diseases and may have implications for improving diagnosis and treatment planning in clinical settings.