Artificially-generated consolidations and balanced augmentation increase performance of U-net for lung parenchyma segmentation on MR images
Fig 6
Impact on functional lung values caused by the improvements in segmentation performance: (a) shows the Perfusion (Q) and Ventilation (V) maps based on the segmentation from CNNBal/NoCons: trained with balanced augmentation and without artificially-generated consolidations. (b) shows Q and V maps based on the segmentation from CNNBal/Cons (proposed method): trained with balanced augmentation with artificially-generated consolidations. Notice, that ventilation and perfusion parameters change essentially when whole lung parenchyma including the real consolidation is correctly segmented using CNNBal/Cons. Ventilation and perfusion defect percentage maps increased substantially (from 20% up to 42% for ventilation and from 12% to 29% for perfusion) with the proposed CNN network. The images were created with Phase Resolved Functional Lung MRI3 (PREFUL) method.