Lung Segmentation on High-Resolution Computerized Tomography Images Using Deep Learning: A Preliminary Step for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Idiopathic pulmonary fibrosis (IPF) requires accurate lung parenchyma segmentation for radiomics studies.
• Manual segmentation is time-consuming and operator-dependent, impacting study reproducibility.
• Developing automated, accurate, and fast segmentation methods is crucial for IPF research.

Purpose of the Study:

• To identify an automatic, accurate, and fast deep learning segmentation approach for lung parenchyma in high-resolution computed tomography (HRCT) images of IPF patients.
• To enhance radiomics studies by providing operator-independent segmentation for texture-based prediction models.
• To evaluate deep learning models on a small dataset of HRCT images.

Main Methods:

• Investigated two deep learning models: U-Net and E-Net.
• Utilized a small dataset of 42 IPF patient HRCT studies (32 for training).
• Compared model performance based on segmentation accuracy (Dice similarity coefficient) and resource requirements.

Main Results:

• E-Net achieved accurate (Dice similarity coefficient = 95.90%), fast (20.32 s), and clinically acceptable lung region segmentation.
• Demonstrated the feasibility of using deep learning for rapid and precise parenchyma segmentation.

Conclusions:

• Deep learning models, specifically E-Net, can efficiently segment and quantify IPF lung parenchyma.
• Automated segmentation provides user-independent results, crucial for radiomics and predictive modeling in IPF.
• This approach eliminates the need for radiologist supervision in segmentation tasks.