J Astley*, A Biancardi, P Hughes, L Smith, H Marshall, J Eaden, N Weatherley, G Collier, J Wild, B Tahir, The University of Sheffield, Sheffield, United Kingdom

• 
  J Astley

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

WE-IePD-TRACK 3-6 (Wednesday, 7/28/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: Deep learning (DL) has shown competency in numerous image segmentation tasks, including delineating ventilated lung volumes from hyperpolarized gas MRI. We previously demonstrated the utility of a VNet convolutional neural network (CNN), trained on a combination of ³He and ¹²⁹Xe scans, in producing segmentations that outperform conventional methods. Here, we compare the performance of several 3D CNNs with varying training strategies for segmenting ventilated lungs on a large multi-nuclear hyperpolarised gas MRI dataset comprising healthy subjects and patients with a range of pulmonary pathologies.

Methods: The dataset contained 743 volumetric hyperpolarized gas MRI scans, with either ³He (248 scans) or ¹²⁹Xe (495 scans) and corresponding expert segmentations, from 326 healthy subjects and patients with pulmonary pathologies. Five experiments were performed to train each CNN: (1,2) the model was trained on either ¹²⁹Xe or ³He images; (3,4) transfer learning was applied to the pre-trained models in (1,2) to fine-tune the network for the opposite gas images; (5) the model was trained on combined ³He and ¹²⁹Xe data. Experiments were performed on the UNet and VNet architectures with cross-entropy loss and PRELU activation functions. Each trained model was evaluated on a combined testing dataset of ³He and ¹²⁹Xe images (n=75). Segmentation accuracy was evaluated via Dice similarity coefficients (DSCs), average boundary Hausdorff distances (Avg-HDs) and volume comparisons using Pearson correlation and Bland-Altman analysis.

Results: The combined UNet model exhibited statistically significant improvements over the combined VNet model using DSC and Avg-HD metrics. Using Pearson correlation and Bland-Altman analysis, DL segmentation volume was highly correlated with the expert segmentations and exhibited minimal bias.

Conclusion: We compared the performance of several 3D CNNs and training strategies for segmenting ventilated lungs on a large multi-nuclear hyperpolarised gas MRI dataset. The UNet CNN trained on both ³He and ¹²⁹Xe data was the highest performing model.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by Yorkshire Cancer Research, Weston Park Cancer Charity, National Institute of Health Research and the Medical Research Council.

ePosters

Keywords

Segmentation, Convolution, Lung

Taxonomy

IM- MRI : Hyperpolarized Imaging

Contact Email