Validation of CT parenchymal and vascular metrics in COPD: outcome-deep, exercise-shallow, and texture-dependent

Question: How have CT-derived spatial and texture-based metrics of lung parenchymal and vascular architecture been validated against pulmonary function, exercise capacity, and mortality in chronic obstructive pulmonary disease cohorts?

Answer

Across the retrieved literature, validation of CT architecture metrics in COPD is strongly asymmetric along three axes. Against pulmonary function, validation is mature but consistently modest in magnitude: the best texture-based classifier of emphysema and functional small airways disease reaches only r ≈ −0.5 with FEV1, and automated mucus plug burden ρ = −0.41, in cohorts of 8,000–9,000 COPDGene participants [8][13]. Against mortality, validation is genuinely incremental — hazard ratios persist after adjustment for emphysema, vessel volume, and scanner — but rests on a small number of single-preprint analyses, one of which reports distance-based vascular coverage as an independent predictor with a 38% mortality excess in its lowest quartile [4][13]. Against exercise capacity, validation is the weakest: it is operationalised almost exclusively as six-minute walk distance, reported as small continuous effect sizes without discrimination, calibration, or reclassification statistics [4][13]. Taken together, these papers support the position that the metric family which has actually cleared outcome validation is spatial-geometric (distance/coverage and plug-count constructs) rather than gray-level textural, while texture-based metrics outperform density-based ones only on the physiology axis and only before acquisition-pipeline effects are accounted for [8][1].

Spirometric calibration is reproducible but consistently explains less than half the variance

Feature-based parametric response mapping using wavelet scattering networks achieved stronger FEV1 correlations than conventional density-based PRM in 8,972 tobacco-exposed COPDGene participants (emphysema r = −0.54, fSAD r = −0.51 versus r = −0.42 for both with standard PRM, P < 0.0001), alongside 95% voxel-wise agreement with the conventional method [8]. AI-quantified airway mucus plug burden in 8,971 COPDGene participants correlated with post-bronchodilator FEV1 % predicted at ρ = −0.41 and with air trapping at ρ = 0.33 [13]. Vascular Proximity Coverage declined monotonically across GOLD strata (5.82 ± 0.50 in non-smokers to 5.21 ± 0.64 in GOLD 4, p < 0.001) and was already lower in current versus former smokers with normal spirometry (5.31 vs 5.68, p < 0.001), demonstrating sensitivity to disease before functional impairment [4]. A lightweight CNN applied to minimally processed CT discriminated the spirometric COPD definition with AUC 0.92 internally, 0.92 in SPIROMICS and 0.82 in NLST, with Brier scores of 0.11–0.17 [12]. The consistent ceiling of r ≈ 0.4–0.55 across structurally different metrics indicates that spirometry is a shared, reproducible but low-specificity validation target — CT architecture measures something spirometry does not capture, but spirometry cannot arbitrate between competing CT metrics.

Mortality validation now tests incremental value, not mere association

The strongest evidence for independence comes from vascular architecture: each SD decrease in VPC was associated with 38% higher mortality in the lowest versus reference quartile (HR 1.38, 95% CI 1.18–1.62) after adjustment for demographics, smoking, emphysema, scanner type and BV5 [4]. Critically, a discordant phenotype with preserved BV5 but low VPC carried 24% higher mortality than the concordant high-BV5/high-VPC group, which the authors frame as evidence that compensatory central dilation can mask peripheral pruning when only total vessel volume is measured [4]. Mucus plug presence was independently associated with all-cause mortality (adjusted HR 1.28) and exacerbation frequency (adjusted IRR 1.32) among GOLD 1–4 participants, with associations reported for respiratory mortality across GOLD categories and cardiovascular mortality in GOLD 1–2 [13]. A separate multi-organ chest-CT aging framework, locked on CT-RATE and MIDRC and then applied to 35,293 participants in NLST and COPDGene, found that organ-specific biological age gaps including thoracic organs tracked all-cause and cause-specific mortality [3]. These designs — phenotyping by discordance from an established metric and locking before external application — are the methodologically strongest validations retrieved.

Exercise capacity is validated only as a continuous 6MWD coefficient

VPC was associated with 6MWD at β = −16.8 m per SD decrease (95% CI −19.5 to −14.0), with SGRQ β = 1.9 and mMRC RR = 1.05 [4]. Mucus plug burden showed ρ = −0.26 with 6MWD and ρ = 0.31 with SGRQ [13]. No retrieved paper validated any CT architecture metric against cardiopulmonary exercise testing, peak oxygen uptake, or ventilatory efficiency. This gap matters because functional capacity itself is strongly prognostic independent of imaging: in a 401-patient prospective COPD cohort, classifying patients by combined physical activity and 6-minute walk test performance produced mortality hazard ratios of 2.37 and 2.87 for the two "cannot" quadrants versus the "can/do" reference [7]. Without adjustment for or comparison against such functional phenotyping, the reported 6MWD coefficients cannot be interpreted as demonstrating that CT adds explanatory power for exercise limitation.

Spatial distribution, not gray-level texture, is where independent prognostic signal has been shown

The only metric with both multivariable-adjusted mortality and lung function evidence is explicitly geometric: VPC integrates the area under a cumulative tissue-coverage curve at 0–10 mm distances from segmented vessel voxels, designed precisely because conventional pruning metrics ignore spatial heterogeneity [4]. By contrast, the texture-based PRMD framework is validated against voxel-wise agreement, FEV1 correlation, noise robustness, and misregistration sensitivity — but not against mortality [8]. The tiered elastic PRM scoring system, which stratifies local tissue to separate reversible small airway disease from progressive destruction, is likewise presented as a characterisation tool without reported outcome validation [5]. This division suggests that "texture" as a descriptor family has been validated where the target is contemporaneous physiology, while "texture" as a prognostic biomarker has been validated only when the construct encodes spatial arrangement relative to anatomy.

Texture descriptors are the least portable component of the pipeline

Ultra-high-resolution photon-counting CT phantoms showed that deep learning versus hybrid iterative reconstruction shifts attenuation histogram modes and GLCM-derived parenchymal texture alongside noise reductions of 7.8% at 6 mGy and 15.2% at 1 mGy, even while lesion contrast remained constant — meaning texture differences can arise from reconstruction alone [1]. Voxel-spacing-aware radiomics demonstrated that neighborhood-sensitive and gradient-derived descriptors are the most affected by preprocessing choices, with isotropic resampling degrading agreement relative to native extraction (median ICC 0.9976 for spacing-aware versus lower agreement for resampling) in 685 LIDC-IDRI nodules [9]. A cystic lung phantom revealed systematic underestimation of individual cyst volumes that increased as cyst size decreased, a bias attributed to partial volume effects and previously unreported [15]. Long-term phantom testing of a clinical photon-counting system over two years showed HU relative error of 0.11–0.30% with noise changes of 1–8 HU tracking software updates, establishing that even HU — the substrate of every density and texture metric — drifts with system changes [14]. A systematic review of 37 datasets found reported human airway wall thickness estimates mutually inconsistent, with only one cluster per subject category yielding anatomically and physiologically plausible values, and imaging modality only weakly explaining the spread [6].

Conflicts and gaps

The central conflict is whether texture outperforms density for physiology or merely for the specific pipeline tested. PRMD's advantage over PRM (r = −0.54 vs −0.42) was demonstrated in a single cohort using one texture engine, and its reported noise robustness (error < 5% versus ~15% overestimation for standard PRM) is itself an implicit admission that the comparator's apparent performance is reconstruction-dependent [8][1]. No retrieved paper pitted a texture metric against a density metric against a spatial-geometric metric on the same mortality endpoint. Second, the VPC finding that BV5 can be preserved while mortality rises challenges the adequacy of all vessel-volume-based validation, yet the evidence is a single retrospective preprint whose abstract still carries bracketed placeholders for study phase and follow-up duration, and no external cohort replication is reported [4]. Third, deep learning COPD detection is validated against the same spirometric criterion used to define the label (FEV1/FVC < 0.70), a circularity that AUC 0.92 cannot resolve and that leaves open whether the model detects emphysema, airway disease, or body habitus [12]. Fourth, vascular architecture metrics with survival validation exist in interstitial disease — intrapulmonary vein enlargement independently associated with worsened survival across three IPF cohorts — but the equivalent venous-side quantification has not been reported in COPD [10]. Fifth, computational vascular compromise modelling has been validated against transplant-free survival and 6MWD in pediatric pulmonary arterial hypertension but not in COPD [11]. Finally, no paper reported whether any CT architecture metric improves risk prediction over FEV1 and emphysema using reclassification, calibration, or decision-curve metrics — the only calibration statistics retrieved were for COPD classification, not for outcomes [12].

What would change the answer

A pre-registered head-to-head comparison in COPDGene or SPIROMICS with all-cause mortality as the primary endpoint, entering density, texture, and spatial-distance metrics simultaneously into one locked multivariable model with net reclassification and decision-curve analysis against FEV1 plus emphysema, would settle whether texture adds prognostic information or only physiological correlation. Same-day repeat CT and multi-vendor matched-dose sub-studies reporting intraclass correlation coefficients for parenchymal GLCM and distance-based features — extending the phantom methodology already applied to nodules to parenchyma [2][1] — would establish whether the reported associations survive measurement error. Replacing 6MWD with cardiopulmonary exercise testing endpoints (peak oxygen uptake, ventilatory efficiency) in a cohort with quantitative CT would test the exercise-capacity axis properly. External and temporal replication of VPC and mucus plug burden in an independent cohort such as ECLIPSE or UK Biobank, with complete reporting of units and follow-up, is the minimum for treating vascular spatial metrics as established. Anchoring distance-based vascular coverage to explant micro-CT and histology, following the multiscale approach used to link venous enlargement to endothelial phenotype in fibrotic lung disease [10], would test whether the biomarker measures the architecture it claims to measure rather than a correlate of emphysema severity.

References

[1] Mei, K., Roshkovan, L., Liu, L. P., Halliburton, S. S., Sharma, S., Ross, S., Stroud, T., Akino, N., Yu, Z., Thompson, R., Dhanaliwala, A. H., Litt, H. I., Noel, P. B.. "Deep Learning Reconstruction for Ultra-High-Resolution Photon-Counting CT of the Lung: Image Quality and Texture Characterization". medRxiv. 2026-09-09T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.04.26361840

[2] Mei, K., Roshkovan, L., Halliburton, S. S., Sharma, S., Ross, S., Yu, Z., Thompson, R., Liu, L. P., Dhanaliwala, A. H., Litt, H., Noel, P. B.. "Evaluation of a CZT-based photon-counting detector CT prototype for low-dose lung cancer screening using patient-specific lung phantoms". medRxiv. 2025-12-31T00:00:00.000Z. URL: https://doi.org/10.64898/2025.12.30.25343218

[3] Sato, J., Salehjahromi, M., Zafar, A., Muneer, A., Xu, X., Zhu, E., Vokes, N. I., Cascone, T., Le, X., Altan, M., Gardner, E. E., Sheshadri, A., Ostrin, E. J., Salahudeen, A. A., Li, T., Merad, M., Chaudhuri, A. A., Gerber, D. E., Kay, F. U., Godoy, M. C. B., Carter, B. W., Shroff, G. S., Byers, L. A., Chung, C., Jaffray, D., Rice, D., Liao, Z., Chang, J. Y., Vaporciyan, A. A., Gibbons, D. L., Wu, C. C., Heymach, J. V., Zhang, J., Wu, J.. "Multi-organ aging quantified from routine chest CT predicts chronic disease risk and mortality". medRxiv. 2026-08-31T00:00:00.000Z. URL: https://doi.org/10.64898/2026.08.26.26361434

[4] Aguilera, S. M., Sthanam, V. L., Nakhmani, A., Dransfield, M. T., Bhatt, S. P., Bodduluri, S.. "Vascular Proximity Coverage: A Novel CT-Based Marker of Pulmonary Vascular Distance to the Lung Parenchyma". medRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.15.26365547

[5] Labaki, W. W., Ram, S., Namvar, A., Bell, A. J., Hoff, B. A., Kazerooni, E. A., Galban, S., Martinez, F. J., Hatt, C. R., Murray, S., Mirkes, E. M., Gorban, A. N., Zinovyev, A., Han, M. K., Galban, C. J.. "Quantitative CT Scoring for Local COPD Severity". medRxiv. 2025-06-09T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.09.25324951

[6] Ola, M. K., Seal, M., Sarwar, R. A., Mattireddy, S. K., Brightling, C. E., Burrowes, K., Kaul, H.. "Reported estimates of human airway dimensions are inconsistent across studies". bioRxiv. 2026-01-20T00:00:00.000Z. URL: https://doi.org/10.64898/2026.01.19.699966

[7] Esteban, C., Barrio, I., Aramburu, A., Monsalve-Cobis, A., Chasco, L., Moraza, J., Najera-Zuloaga, J., Aizpiri, S., Aburto, M., Arostegui, I., Quintana, J. M.. "Functional Capacity and Physical Activity in COPD: Associations with Hospitalizations, Health related Quality of Life and Mortality. A Prospective study". medRxiv. 2026-01-06T00:00:00.000Z. URL: https://doi.org/10.64898/2026.01.04.26343342

[8] Namvar, A., Shan, B., Hoff, B., Labaki, W. W., Murray, S., Bell, A. J., Galban, S., Kazerooni, E. A., Martinez, F. J., Hatt, C. R., Han, M. K., Galban, C. J., Ram, S.. "Feature-Based Parametric Response Mapping on Thoracic Computed Tomography for Robust Disease Classification in COPD". medRxiv. 2026-04-27T00:00:00.000Z. URL: https://doi.org/10.64898/2026.04.24.26351675

[9] David Corral Fontecha, Juan Miranda Bautista, Pablo Menendez Fernández-Miranda, Sergio Rubio-Martín, Lara Lloret Iglesias, Jose A. Vega. "Physically Aware Radiomics Without Interpolation: Disentangling Voxel Geometry and Signal Modification in CT and MRI". arxiv. 2026-09-15T00:00:00.000Z. URL: https://arxiv.org/abs/2607.12399

[10] Yamada, D., Wang, H., Szmul, A., Riese, J., Egashira, R., van Moorsel, C. H. M., Mogulkoc, N., Gokce, B., Aslani, S., Cheng, D., Chen, Y., Bodenmann, R., Brunner, M., Tielemans, B., Brunet, J., Urban, T., Purzycka, J., Dejea, H., Tafforeau, P., Vermant, M., van Beek, F. T., Maurer, B., Krattli, S., Kreuzer, M., Bruggmann, R., Follet, T., Veltkamp, M., Goos, T., Anseele, K., Vander Kuylen, M., van Es, H. W., Engel, K., Savas, R., Brenes, J. M., Hendriks, J. M., Lapperre, T. S., Walsh, C. L., Jonigk, D., Neubert, L., Werlein, C., Freitag, H., Dettmer, S., Schupp, J. C., de Sadeleer, L. J., Robin. "Multiscale mapping of venous remodelling in idiopathic pulmonary fibrosis". bioRxiv. 2026-10-02T00:00:00.000Z. URL: https://doi.org/10.64898/2026.09.17.752253

[11] Niccum, M., Avitabile, C., Albizem, D., Meluskey, H., penney, c., Hanna, B. D., O'Byrne, M. L., Bshouty, Z., Frank, D. B.. "Pulmonary Vascular Compromise is Associated with Survival in Pediatric Pulmonary Hypertension: A New Computational Model". medRxiv. 2025-03-26T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.25.25324654

[12] Rabby, A. S. A., Chaudhary, M. F. A., Saha, P., Sthanam, V., Nakhmani, A., Zhang, C., Barr, R. G., Bon, J., Cooper, C. B., Curtis, J. L., Hoffman, E. A., Paine, R., Puliyakote, A. K., Schroeder, J. D., Sieren, J. C., Smith, B., Woodruff, P. G., Reinhardt, J. M., Bhatt, S. P., Bodduluri, S.. "Light Convolutional Neural Network to Detect Chronic Obstructive Pulmonary Disease (COPDxNet): A Multicenter Model Development and External Validation Study". medRxiv. 2025-08-01T00:00:00.000Z. URL: https://doi.org/10.1101/2025.07.30.25332459

[13] Oyer, J., Namvar, A., Hoff, B. A., Bosma, C., Labaki, W. L., Kazerooni, E. A., Martinez, F. J., Hatt, C. R., Han, M. K., Galban, C. J., Ram, S.. "Artificial Intelligence-Based Detection of Airway Mucus Plugs on CT and Associations With Clinical Outcomes in COPDGene". medRxiv. 2026-06-15T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.10.26355393

[14] Liu, L. P., Pasyar, P., Liu, F., Cao, Q., Sandvold, O. F., Rybertt, M. V., Sahbaee, P., Shinohara, R. T., Litt, H. I., Noël, P. B.. "Assessing the Stability of Photon-Counting CT: Insights from a Two-Year Longitudinal Study". medRxiv. 2024-12-14T00:00:00.000Z. URL: https://doi.org/10.1101/2024.06.05.24308046

[15] Shefra Shah, Farah Hussaini, Dumitru Mazilu, Eric E. Bennett, Han Wen. "Cystic Lung Phantom to Validate Clinical CT Protocols". arxiv. 2025-03-26T00:00:00.000Z. URL: https://arxiv.org/abs/2503.19747

How this survey was made

This survey answers one question: “How have CT-derived spatial and texture-based metrics of lung parenchymal and vascular architecture been validated against pulmonary function, exercise capacity, and mortality in chronic obstructive pulmonary disease cohorts?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 15 papers whose work this text draws on, chosen from the 123 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 22, 2026 and covered work from the previous two years.