Data-Efficient Lung Segmentation Using Foundational Models: Improving Clinical Workflow With Segment Anything Model (SAM) for Hyperpolarized Gas MRI

Abstract Introduction: Hyperpolarized 129Xe/3 He lung MRI is a well-established method for evaluating pulmonary conditions and ventilation defects [1,2]. Traditional segmentation techniques [3] are effective but labor-intensive and time-consuming. Despite advances in deep learning (DL)-based segmentation of hyperpolarized gas MRI [4-6], the use of foundational models in this domain has been underutilized [7]. Our proof-of-concept research hypothesis is that foundational models, such as the Segment Anything Model (SAM) [8], can be utilized as efficient, ready-to-deploy for lung MRI segmentation in data-limited clinical environments. We propose that SAM can achieve superior segmentation performance compared to traditional convolutional neural networks (CNNs) even when trained on small datasets, making it a clinically viable tool for real-world use. Methods: We collected data from 56 participants (9 healthy, 28 COPD, 9 asthma, 10 COVID-19), resulting in 896 2D slices (128x128 pixels). The dataset was split 80% for training and 20% for testing. To assess data efficiency, experiments were performed with 25% of the training data. SAM was fine-tuned for segmentation using PyTorch on an NVIDIA GA102 GPU and compared to CNN-based models (UNet with ResNet18, ResNet152, VGG16, and VGG19). All models were evaluated on the same test set for consistency. Results: With only 25% of the available training data, SAM achieved a DSC of 0.97 for both proton and hyperpolarized gas MRI, significantly outperforming CNN models. UNet with VGG16 scored 0.89 for proton MRI and 0.78 for hyperpolarized gas MRI, while UNet with ResNet18 achieved 0.89 and 0.84, respectively. These results suggest that SAM can provide near-expert-level performance even in data-limited clinical settings. Conclusion: SAM outperformed traditional CNNs, particularly in data-limited conditions, demonstrating its potential as a robust, ready-to-deploy segmentation tool in clinical practice. Its ability to achieve high performance with minimal training data suggests SAM could be implemented in clinics where large datasets are not feasible, offering clinicians a fast, accurate, and scalable solution for lung MRI segmentation. Future work will explore one-shot learning techniques and larger datasets to enhance clinical applicability further. References: 1_Perron, S. et al. J_Magn_Reson_348, 107387 (2023).2_Kirby, M. et al. J_Appl_Physiol_114, 707-715 (2013). 3_Kirby, M. et al. Acad_Radiol_19, 141-152 (2012). 4_Astley, J. R. et al. Sci_Rep_12, 10566 (2022). 5_Astley, J. R. et al. J_Magn_Reson_Imaging_57, 1878-1890 (2022). 6_Astley, J. R. et al. Br_J_Radiol_95 (2022). 7_Babaeipour, R. et al. Bioengineering_10, 1349 (2023). 8_Kirillov, A. et al. ICCV, 3992-4003(2023).

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.