Lung Segmentation on HRCT and Volumetric CT for Diffuse Interstitial Lung Disease Using Deep Convolutional Neural...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate lung segmentation is crucial for diagnosing diffuse interstitial lung disease (DILD).
• Conventional methods may struggle with the complex patterns seen in DILD across different CT protocols.
• Deep learning offers potential for improved automated segmentation.

Purpose of the Study:

• To develop and evaluate a robust deep convolutional neural network (CNN) based lung segmentation method for DILD.
• To assess the performance of the U-Net model on both high-resolution computed tomography (HRCT) and volumetric CT datasets.
• To compare the deep learning method against conventional segmentation techniques.

Main Methods:

• A 2D U-Net deep CNN architecture was trained on HRCT images from 617 DILD patients.
• Manual segmentation by a radiologist served as the gold standard.
• The model was further evaluated on 30 independent volumetric CT scans of usual interstitial pneumonia (UIP) patients.
• Segmentation accuracy was quantitatively assessed using Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), mean surface distance (MSD), and Hausdorff surface distance (HSD).

Main Results:

• The U-Net based deep learning method demonstrated significantly superior lung segmentation performance compared to conventional methods (p < 0.001).
• High accuracy metrics were achieved on HRCT images: DSC 98.84±0.55%, JSC 97.79±1.07%, MSD 0.27±0.18 mm, HSD 25.47±13.63 mm.
• Segmentation accuracy for volumetric CT scans was comparable to that of HRCT scans.

Conclusions:

• A highly accurate and robust U-Net based lung segmentation method for DILD has been developed.
• The method is effective across different CT imaging protocols, including HRCT and volumetric CT.
• This deep learning approach holds promise for improving the analysis and diagnosis of DILD.