Implementable Deep Learning for Multi-sequence Proton MRI Lung Segmentation: A Multi-center, Multi-vendor, and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning, specifically convolutional neural networks (CNNs), has become the standard for proton (¹H)-MRI lung segmentation.
• Previous CNN studies were limited by single-center data and restricted acquisition parameters.

Purpose of the Study:

• To develop a generalizable CNN for ¹H-MRI lung segmentation.
• The model aims for robustness against various pathologies, acquisition protocols, vendors, and imaging centers.

Main Methods:

• A retrospective study utilized 809 ¹H-MRI scans from 258 participants with pulmonary pathologies and 31 healthy participants.
• 2D and 3D CNNs were trained and compared against the spatial fuzzy c-means (SFCM) method and manual segmentations.
• Performance was assessed using Dice similarity coefficient (DSC), average boundary Hausdorff distance (Average HD), and relative error (XOR) on testing and external validation datasets.

Main Results:

• The 3D CNN significantly outperformed the 2D CNN and SFCM.
• On the testing set, the 3D CNN achieved a median DSC of 0.961, Average HD of 1.63 mm, and XOR of 0.079.
• On external validation data, the 3D CNN yielded a median DSC of 0.973, Average HD of 1.11 mm, and XOR of 0.054.

Conclusions:

• The developed 3D CNN accurately segments lungs in ¹H-MRI.
• The model demonstrates significant robustness to variations in disease pathology, MRI sequence, vendor, and imaging center.