Comparative Analysis of Foundational, Advanced, and Traditional Deep Learning Models for Hyperpolarized Gas MRI Lung...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Chronic obstructive pulmonary disease (COPD) necessitates advanced diagnostic tools.
• Hyperpolarized gas MRI (using ³He and ¹²⁹Xe) provides non-invasive lung function assessment.
• Accurate segmentation is vital for analyzing lung MRI data.

Purpose of the Study:

• To compare the performance of foundational, advanced, and traditional deep learning models for hyperpolarized gas MRI segmentation.
• To evaluate model performance under varying data availability scenarios (100% down to 10%).
• To identify robust segmentation strategies for data-scarce medical imaging applications.

Main Methods:

• Comparative analysis of Segment Anything Model, MedSAM, UniRepLKNet, TransXNet, UNet, FPN, and DeepLabV3.
• Segmentation performance evaluated using Dice Similarity Coefficient (DSC) across four data reduction levels.
• Statistical significance assessed using p-values for performance comparisons.

Main Results:

• Foundational and advanced models demonstrated statistically equivalent performance across all data scenarios (p > 0.01).
• Both foundational and advanced models significantly outperformed traditional models under data constraints (p < 0.001).
• At 10% training data, foundational/advanced models maintained DSC > 0.86, while traditional models failed.

Conclusions:

• Architectures with large effective receptive fields are crucial for medical imaging segmentation with limited data.
• Foundational and advanced models offer a promising solution for democratizing advanced medical imaging analysis in resource-limited settings.
• This study underscores the adaptability and robustness of newer AI architectures in challenging clinical data scenarios.