Comparative Analysis of Foundational and Traditional Deep Learning Models for Hyperpolarized Gas MRI Lung Segmentation: Robust Performance in Data-Constrained Scenarios

This study investigates the comparative performance of foundational models versus traditional deep learning architectures for hyperpolarized gas MRI segmentation under both full data and limited data conditions. Chronic obstructive pulmonary disease (COPD) remains a leading global health concern, and advanced imaging techniques are crucial for its diagnosis and management. Hyperpolarized gas MRI, utilizing helium-3 (³He) and xenon-129 (¹²⁹Xe), offers a non-invasive way to assess lung function. Foundational models, pre-trained on diverse and expansive datasets, theoretically offer advantages in scenarios with limited task-specific data compared to traditional architectures that rely heavily on large training datasets. This study evaluates this hypothesis by comparing foundational models, Segment Anything Model (SAM) and Segment Anything in Medical Images (MedSAM) against traditional deep learning architectures (UNet with VGG19 backbone, Feature Pyramid Network with MIT-B5 backbone, and DeepLabV3 with ResNet152 back-bone) using both full dataset (1640 2D MRI slices from 205 participants) and limited data scenarios (25% of the full dataset). Our experimental design included fine-tuning all models on the complete dataset and subsequently on the reduced dataset to assess performance degradation and resilience to data scarcity. Results demonstrate that while traditional models achieve competitive performance with full data availability, foundational models exhibit superior robustness and maintain performance in limited data scenarios. The fine-tuned MedSAM model showed the least performance degradation when transitioning from full to limited data conditions, achieving superior Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD95) metrics compared to traditional architectures. This work highlights the critical advantage of foundational models in medical imaging applications where data collection is challenging, expensive, or ethically constrained, demonstrating their potential to democratize advanced medical imaging analysis in resource-limited settings.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.