Key breakthroughs are required to advance novel imaging technologies with respect to research and development, regulatory and patient safety issues and clinical practicability. Here we focus on clinical translation of pulmonary MRI – structure-function methods needed for clinical use. In order to enable clinical translation of pulmonary functional imaging for large-scale, multi-centre and longitudinal applications, we think it is critical to develop image analysis tools that can provide clinically-acceptable and clinically-relevant measurements of regional anatomical and functional abnormalities while integrated into and enhancing clinical workflows. This is important because pulmonary image acquisition is relatively rapid, but image analysis is complex involving multiple modality registration, segmentation and biomarker quantification. Therefore, the objective of this work was to develop a pulmonary segmentation pipeline for quantitative evaluation of ¹²⁹Xe MRI ventilation across a range of pulmonary abnormalities.

Subjects and image acquisition:

In total, 20 subjects including eight asthmatics, seven COPD and five healthy subjects provided written informed consent to a study protocol approved by Health Canada. MRI was performed using a whole body 3.0T Discovery MR750 system (General Electric Health Care, Milwaukee, Wisconsin, USA). ¹H MRI was performed using a whole-body radio-frequency coil and a ¹H fast spoiled gradient-recalled echo (FGRE) sequence using a partial echo (16s acquisition time; repetition time (TR)/echo time (TE)/flip angle=4.3ms/1.2ms/20°; field-of-view (FOV)=40cm×40cm; matrix=128×80; number of slices=14–17; slice thickness=15mm). Hyperpolarized ¹²⁹Xe MRI static ventilation imaging was performed using a 3D FGRE sequence (acquisition time=16s; TR/TE/flip-angle=7.0ms/1.8ms/variable; FOV=40cm×40cm; matrix=128×128; BW=9kHz; NEX=1; number of slices=16; slice thickness=15mm) during a breath-hold.

¹H-¹²⁹Xe image segmentation:

¹H and ¹²⁹Xe images were resampled to ~3 mm isotropic space using Convert3D¹ tools prior to algorithm segmentation. One observer placed seeds on the left and right lung and background on the resampled ¹H and ¹²⁹Xe images three times on 3 different days. A rigid, deformable registration approach² was used to register the resampled ¹²⁹Xe and ¹H images and the derived deformation field was used to deform the ¹²⁹Xe seeds to the ¹H image space. These seeds were used to estimate the appearance models for the respective regions and jointly formulate the data terms of the Potts approach³ while the regularization terms were generated based on ¹H MRI edge information only. The resultant non-convex optimization problem involving binary labeling functions for the three regions was approximated through convex relaxation and further represented by an equivalent max-flow model through variational analysis. ¹²⁹Xe images were then hierarchically segmented to generate ventilation defects percent (VDP).

Validation:

The accuracy of ¹H-¹²⁹Xe segmentation was evaluated by comparing algorithm lung masks with manual results performed by an experienced observer. We employed the Dice similarity coefficient (DSC), root-mean-squared-error (RMSE) of the Euclidian distance between the two segmented lung surfaces and absolute volume error (|dVe|) to measure the agreement of the two sets of lung volumes. We also compared VDP generated using this new method with a semi-automated approach, previously described.⁴ Moreover, reproducibility was measured by calculating coefficient of variability (CoV) and intra-class correlation coefficient (ICC) of DSC, RMSE and VDP. The computational efficiency was obtained by averaging the runtime of the repeated algorithm segmentation.

Representative ¹H-¹²⁹Xe segmentation results are shown in Figure 1. The proposed approach yielded a DSC of 89 ±2%, RMSE of 4±1mm and |dVe| of 0.50±0.40L for whole lung. The CoV(ICC) were 0.32%(0.99), 2.4%(0.98), 4%(1.00) for DSC, RMSE and VDP, respectively. Pearson correlation results showed that VDP for five healthy and seven COPD subjects was significantly and strongly correlated with the semi-automated measurements (Figure. 2) with no statistically significant difference (p=0.14). The mean runtime for the proposed approach was ~1 min for each subject, including ~15s for seeding, ~20s for registration, ~5s for max-flow segmentation and ~15s for ventilation defect/VDP generation.The proposed ¹H-¹²⁹Xe segmentation pipeline generates lung volumes and VDP that are in agreement with manual and semi-automated reference standard. This approach involves diminished user interaction and rapid implementation compared to a runtime of ~15min for the semi-automated method. Importantly, the proposed pipeline is highly reproducible in generating lung volumes and providing VDP (CoV of ~4% vs ~20% by the semi-automated approach) from pulmonary ¹H and ¹²⁹Xe MRI. These results suggest that the proposed pipeline might be suitable for large-scale, multi-centre and longitudinal clinical applications.