3560

View Abstract / View Presentation

Comparison of 3D convolutional neural networks and loss functions for ventilated lung segmentation using multi-nuclear hyperpolarized gas MRI

Joshua R Astley^1,2, Alberto M Biancardi¹, Paul J Hughes¹, Laurie J Smith¹, Helen Marshall¹, Guilhem J Collier¹, James Eaden¹, Nicholas D Weatherley¹, Jim M Wild¹, and Bilal A Tahir^1,2
¹POLARIS, University of Sheffield, Sheffield, United Kingdom, ²Oncology and Metabolism, University of Sheffield, Sheffield, United Kingdom

We compared the performance of several 3D convolutional neural network architectures and loss functions for segmentation of ventilated lungs on a large and diverse multi-nuclear hyperpolarized gas MRI dataset.

Figure 3. Example coronal slices of the UNet and VNet combined ³He and ¹²⁹Xe trained segmentations for three cases with different diseases compared to the expert segmentations. DSC and Avg HD values are given for each case.

Figure 4. Correlation and agreement analysis of lung volumes for 75 testing set cases compared to expert segmentations for combined ³He and ¹²⁹Xe a) VNet and b) UNet models.