Using 3D point cloud and graph-based neural networks to improve the estimation of pulmonary function tests from chest...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Pulmonary function tests (PFTs) are crucial for assessing interstitial lung disease in systemic sclerosis, but spirometry has limitations.
• Previous deep learning models (CNN-CT, CNN-Vessel) estimated PFTs from CT scans but lost detail due to image down-sampling.
• The impact of detailed lung vessel structure on PFTs remains incompletely understood.

Purpose of the Study:

• To develop novel deep learning models (PNN-Vessel, GNN-Vessel) for estimating PFTs using detailed pulmonary vessel centerline information.
• To compare the performance and efficiency of these new models against existing methods.
• To investigate the contribution of detailed vessel data to PFT estimation when combined with CT scan information.

Main Methods:

• Utilized point cloud neural networks (PNN-Vessel) and graph neural networks (GNN-Vessel) on pulmonary vessel centerlines derived from CT scans.
• Compared PNN-Vessel and GNN-Vessel against a CNN-based approach (CNN-Vessel) using intra-class correlation coefficient (ICC) for four PFT metrics.
• Employed multiple variable stepwise regression to combine CNN-CT with PNN-Vessel and GNN-Vessel for optimal PFT estimation.

Main Results:

• PNN-Vessel and GNN-Vessel significantly outperformed CNN-Vessel, with average ICC improvements of 14% and 4%, respectively.
• PNN-Vessel and GNN-Vessel demonstrated superior efficiency, using less training time and fewer parameters compared to CNN-Vessel.
• The combined model (CNN-CT + PNN-Vessel + GNN-Vessel) achieved the highest PFT estimation accuracy, with ICCs ranging from 0.742 to 0.836.

Conclusions:

• Detailed pulmonary vessel information, effectively captured by PNN-Vessel and GNN-Vessel, significantly enhances PFT estimation accuracy.
• These novel network architectures offer a more efficient and informative approach to PFT prediction from anatomical imaging.
• The findings support the use of advanced deep learning techniques for non-invasive assessment of lung function in interstitial lung diseases.