Deep convolutional neural networks with multiplane consensus labeling for lung function quantification using UTE...

Area of Science:

• Medical Imaging
• Pulmonary Medicine
• Artificial Intelligence

Background:

• Ultrashort echo time (UTE) proton MRI is valuable for lung imaging.
• Labor-intensive lung segmentation hinders rapid biomarker extraction and regional quantification.
• Developing automated methods is crucial for advancing functional lung imaging analysis.

Purpose of the Study:

• To evaluate a deep learning (DL) approach for automated lung segmentation.
• To extract image-based biomarkers from functional lung imaging using 3D radial UTE oxygen-enhanced (OE) MRI.
• To assess the efficiency and accuracy of DL-based segmentation compared to a reference method.

Main Methods:

• A retrospective study evaluated a 2D convolutional encoder-decoder DL method for automated lung segmentation.
• 3D radial UTE (TE = 0.08 msec) sequences were acquired under normoxic and hyperoxic conditions in 45 subjects.
• DL segmentation and functional quantification were compared with supervised region growing using Dice coefficients, Wilcoxon signed-rank tests, and Bland-Altman analysis.

Main Results:

• The DL method demonstrated strong agreement with the reference method for lung segmentation (Dice: 0.97 for right, 0.96 for left).
• DL automated segmentation in an average of 46 seconds, compared to 1.93 hours for the reference method (P < 0.001).
• Bland-Altman analysis revealed nonsignificant inter-method differences in volumetric and functional measurements.

Conclusions:

• Deep learning provides rapid, automated, and robust lung segmentation for UTE proton MRI.
• This approach facilitates efficient quantification of regional lung function.
• DL-based segmentation is a viable and effective tool for pulmonary imaging biomarker extraction.