CT-Based Local Distribution Metric Improves Characterization of COPD

Parametric response mapping (PRM) of paired CT lung images has been shown to improve the phenotyping of COPD by allowing for the visualization and quantification of non-emphysematous air

trapping component, referred to as functional small airways disease (fSAD). Although promising, large variability in the standard method for analyzing PRMfSAD has been observed. We postulate

that representing the 3D PRMfSAD data as a single scalar quantity (relative volume of PRMfSAD) oversimplifies the original 3D data, limiting its potential to detect the subtle progression

of COPD as well as varying subtypes. In this study, we propose a new approach to analyze PRM. Based on topological techniques, we generate 3D maps of local topological features from 3D

PRMfSAD classification maps. We found that the surface area of fSAD (SfSAD) was the most robust and significant independent indicator of clinically meaningful measures of COPD. We also

confirmed by micro-CT of human lung specimens that structural differences are associated with unique SfSAD patterns, and demonstrated longitudinal feature alterations occurred with worsening

pulmonary function independent of an increase in disease extent. These findings suggest that our technique captures additional COPD characteristics, which may provide important

opportunities for improved diagnosis of COPD patients.

Chronic obstructive pulmonary disease (COPD) is a major cause of morbidity, mortality, and healthcare cost worldwide with an estimated global prevalence of approximately 12% of adults aged

≥30 years in 2010 and rising with the aging population1, 2. Recent reports found that COPD etiology varies among populations, including risk factors such as tobacco smoke, cooking fuels,

environmental pollution and family genetics2. This has led to the current understanding that COPD covers a wide spectrum of pathophysiologies3, 4. Clinical presentation and monitoring of

COPD have been described primarily through spirometry as pulmonary function measurements. Although highly reproducible, these measures, such as forced expiratory volume in one second (FEV1),

assess the lungs as a whole and are unable to differentiate two key components of COPD: emphysema and small airways disease. In addition, spirometry does not provide spatial context for

regional heterogeneity of these components. X-ray computed tomography (CT) has addressed some of these limitations by allowing clinicians to verify emphysema in patients exhibiting loss of

pulmonary function. Even with these techniques, COPD is often undiagnosed in early stages, impeding proper treatment with the disease progressing to permanent lung damage (i.e. emphysema).

Although COPD phenotyping has been prolifically reported in the literature5,6,7, lack of accurate diagnostic tools that identify these unique COPD subtypes have hampered the development of

effective therapies. Nevertheless, significant advances in technologies are providing physicians opportunities to shift towards “precision medicine”.

Various strategies have been undertaken to identify metrics that more accurately assess COPD subtypes, such as genetic, molecular and cellular markers as well as medical imaging devices and

methodologies. Although advances in biological phenotyping have shown promise in identifying disease heterogeneity in patients4, 8, these approaches are generally either global measures or

highly invasive. In contrast, medical imaging provides clinicians with a relatively non-invasive and reproducible approach that provides functional information that is spatially defined.

Although various instruments (e.g. PET, SPECT and MRI) are heavily investigated as surrogates of pulmonary function and clinical outcome9, CT, with its high resolution and lung contrast,

continues to be the most widely used medical imaging device in the clinic. As such, advances in this technology would have an immediate impact on patient care.

CT is inherently a quantitative map, where x-ray attenuation is linearly proportional to lung tissue density10, 11. Extensive research in CT image post-processing has generated an array of

potentially diagnostic and prognostic measures. Filter-based techniques and airway wall measurements have been extensively explored12,13,14. Not only have these methodologies advanced our

understanding of COPD, they are also becoming more prevalent in clinic decision-making. In fact, the quantification of discrete phenotypes of emphysema using CT has had an impact on patient

care. At present three emphysema patterns (i.e., centrilobular, panlobular, and paraseptal emphysema) have been identified, each of which are strongly associated with a range of respiratory

physiologies and functional measures15, 16. The understanding that unique spatial patterns of emphysema serve as indicators of COPD subtypes has spawned progress in lobe segmentation

algorithms17, 18 as well as the need to evaluate CT-based features19. Although our understanding related to the clinical implications of the spatial patterns of emphysema is emerging20,

little is understood about the non-emphysematous component of COPD, commonly associated with small airways disease.

Small airway disease, a major component of COPD, is generally characterized by the presence of inflammation, fibrosis, and mucous plugging, all of which contribute to airflow obstruction. At

less than 2 mm in diameter, these airways are essentially invisible to clinical imaging scanners, hindering accurate COPD phenotyping by CT especially when emphysema is radiographically

identified. The Parametric Response Mapping (PRM) technique21 addressed this limitation. Through the spatial alignment of paired inspiration and expiration CT scans, PRM of CT data

delineates and quantifies non-emphysematous air trapping, an indirect measure of small airways disease (SAD), even in the presence of emphysema21, 22. The extent of fSAD, as measured by PRM

as the relative volume (%PRM) in the lungs, has been reported to be an independent indicator of pulmonary function decline as well as other clinically relevant measures, re-affirming

previous histological studies22,23,24. In addition, McDonough and colleagues have shown pathologically in human core lung specimens imaged by micro-CT that small airways disease may in fact

serve as a precursor to emphysema25. This highlights the potential importance of the independent and non-invasive evaluation of fSAD through PRM26. Although the spatial information of fSAD

is retained as a 3D PRM classification binary map, studies have primarily focused on the use of a whole-lung measure of fSAD, presented as a relative lung volume, which serves as the extent

of this COPD component within the patient. As observed with emphysema, the spatial distribution of fSAD may aid treating physicians by providing them unique diagnostics that serve as a

surrogate of clinically meaningful outcomes.

The present study demonstrates an extension of the PRM approach that extracts local topological features from PRM-derived disease classifications maps (Fig. 1). Using CT scans from COPD

patients assessed as GOLD stage 1–4 accrued as part of a well-controlled multi-center clinical trial27, we found that disease pattern (i.e. topological features) is correlated with COPD

severity independent of disease extent (i.e. relative volume). Through micro-CT analysis of explanted lung cores from a lung transplant recipient with bronchiolitis obliterans syndrome, an

obstructive lung disease‚ and longitudinal CT data acquired from a COPD subject, we provide anecdotal evidence that PRM-derived topological features are associated with structural

differences and may also reveal trends in progressing obstructive disease. This work demonstrates for the first time that spatial features extracted from PRMfSAD maps, specifically the

surface area (SfSAD), provide independent predictors of clinical outcome measures, as well as provide illustrative examples that these features are associated with unique airway and

parenchyma structures and disease progression.

A schematic of the workflow is displayed for generating PRM topological maps. (A) CT images are acquired at expiration and inspiration. (B) PRM analysis is performed by first segmenting the

lungs from the thoracic cavity. Then the CT images are filtered and spatially aligned to the expiration geometric frame. Individual voxels are then classified as normal (PRMNorm, green),

functional small airways disease (PRMfSAD, yellow), or emphysema (PRMEmph, red). (C) Topological feature extraction is performed on each PRM classification binary map to determine topology

metrics. Presented are surface area (S) maps for PRMfSAD (left) and PRMEmph (right).

The topological features of the PRM classification binary maps of fSAD and emphysema, defined throughout as PRMfSAD and PRMEmph, were determined using the Minkowski Functionals: surface area

(Si), mean curvature length (Bi), the Euler-Poincare characteristic (χi), and a condensed descriptor of clustering (αi), where i is an index for fSAD or Emph determined by PRM. These

measures were determined locally, referred to as “Local”, over sub-volumes of the lung using a moving window approach resulting in a 3D parameter map for each metric, or globally, referred

to as “Global”, over the entire lung volume resulting in a single parameter scalar quantity for each metric. For statistical analysis Local values represent the full lung mean value. Four

Local and four Global topological metrics were generated from each binary PRM classification map. We first sought to determine the robustness of each parameter by performing a linear

regression of the mean of Local parameters to their Global parameters when applied to PRMfSAD, all voxels classified as fSAD by PRM, and PRMEmph, all voxels classified as emphysema by PRM,

binary maps. We observed for the surface area (S) of PRMfSAD and PRMEmph near perfect agreement between Local and Global calculations (Supplemental Fig. 3). Linear regression of Local and

Global Si generated R2 of >0.999. Increasing complexity of the topology metric was found to demonstrate less correlation between Local and Global calculations with α demonstrating clear

disagreement between measures (R2 −856 HU on expiration), functional small airways disease (PRMfSAD, yellow; >−950 HU on inspiration and ≤−856 HU on expiration), and emphysema (PRMEmph, red;

≤−950 HU on inspiration and ≤−856 HU on expiration). Whole-lung measures from PRM analysis are reported as the relative lung volume for each classification (%PRMi, where i is an index for

fSAD or Emph determined by PRM). In order to minimize the contribution of blood vessels and airways in the analysis, all voxels with HU values >−500 HU in either scan were omitted.

Topological properties of each PRM classification map were explored as independent indicators of clinical outcome (Fig. 1). These topological properties were defined in this study through

the Minkowski measures (local estimates of the Minkowski functionals) associated with 3D distributions: Volume (V, in mm3), Surface Area (S, in mm2), Mean Breadth (B, in mm), and the

Euler-Poincaré statistic (χ)33. Additional processing with use of the χ statistic produced a condensed descriptor of clustering, α (Supplemental Methods). A detailed description of these

parameters is provided in the supplement (Supplemental Methods and Supplemental Fig. 1). Maps of Minkowski measures (i.e. V, S, B, χ and α) were computed using a moving window of size 213

evaluated on a grid of every 5th voxel. Local values from each parameter were normalized to produce parametric densities, with V, S, and B normalized by the masked local window volume and χ

and α were normalized by the masked window voxel count. Minkowski measures were quantified per subject as the mean local normalized value over the entire lung volume for group comparisons

and regression. For display purposes (Figs 1 and 2 and Supplemental Fig. 4), we multiplied Minkowski measures (S, B, χ and α) by the local density, V, to highlight regions of substantial

disease. Final displayed representations of spatially resolved indices have been linearly interpolated back to original dimensions. In addition, global values for V, S, B, χ and α were

calculated for each PRM classification over the entire lung volume (Supplemental Methods). The expected behavior of each metric was evaluated using simulations of random distributions at

each relative volume (Supplemental Methods and Supplemental Fig. 2). Parameter V is analogous to relative volumes of PRM classification. As such, this parameter was not included in the study

analyses. All image processing were performed using in-house algorithms developed in a technical computing language (MATLAB, The MathWorks Inc., Natick, MA).

Differences in metrics between GOLD were assessed by ANOVA using a Bonferroni post-hoc test to account for multiple comparisons. Topology (i.e. Si, Bi, χi and αi) and extent (%PRMi) of

disease were evaluated as independent indicators of various clinical outcomes by multivariate linear or logistic regression analysis with stepwise entry. Regression analysis included age,

gender, and body mass index (BMI) to account for known clinical correlations. All statistical computations were performed using a statistical software package (SPSS Software Products).

Results were considered statistically significant at the two-sided 5% comparison-wise significance level (P