Purpose
To develop an image-processing pipeline for semiautomated (SA) and reproducible analysis of hyperpolarized gas lung ventilation and proton anatomical magnetic resonance imaging (MRI) scan pairs. To compare results from the software for total lung volume (TLV), ventilated volume (VV), and percentage lung ventilated volume (%VV) calculation to the current manual “basic” method and a K-means segmentation method.
Materials and Methods
Six patients were imaged with hyperpolarized 3He and same-breath lung 1H MRI at 1.5T and six other patients were scanned with hyperpolarized 129Xe and separate-breath 1H MRI. One expert observer and two users with experience in lung image segmentation carried out the image analysis. Spearman (R), Intraclass (ICC) correlations, Bland–Altman limits of agreement (LOA), and Dice Similarity Coefficients (DSC) between output lung volumes were calculated.
Results
When comparing values of %VV, agreement between observers improved using the SA method (mean; R = 0.984, ICC = 0.980, LOA = 7.5%) when compared to the basic method (mean; R = 0.863, ICC = 0.873, LOA = 14.2%) nonsignificantly (pR = 0.25, pICC = 0.25, and pLOA = 0.50 respectively). DSC of VV and TLV masks significantly improved (P < 0.01) using the SA method (mean; DSCVV = 0.973, DSCTLV = 0.980) when compared to the basic method (mean; DSCVV = 0.947, DSCTLV = 0.957). K-means systematically overestimated %VV when compared to both basic (mean overestimation = 5.0%) and SA methods (mean overestimation = 9.7%), and had poor agreement with the other methods (mean ICC; K-means vs. basic = 0.685, K-means vs. SA = 0.740).
Conclusion
A semiautomated image processing software was developed that improves interobserver agreement and correlation of lung ventilation volume percentage when compared to the currently used basic method and provides more consistent segmentations than the K-means method.
Level of Evidence: 3
Technical Efficacy: Stage 2
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)