Computed tomography features associated with pneumothorax susceptibility in pectus excavatum: a retrospective case-control study of lobar morphometry, attenuation metrics, and machine-learning projection Authors: In Sub Kim, Jae Jun Kim, Kyong Shil Im, Yoon Ho Kim, Jeong Hwan Ryu Published: 2026 ABSTRACT BackgroundPrimary spontaneous pneumothorax (PSP) can occur without visible blebs or bullae, suggesting a broader structural susceptibility. Pectus excavatum (PE) is associated with a high incidence of PSP. We evaluated chest computed tomography (CT)-based lobar signatures of PSP and tested whether PE exhibits a PSP-like phenotype.MethodsYoung adults aged 20–30 years were retrospectively grouped as first-episode PSP [patients with spontaneous pneumothorax (PT)], PE without prior pneumothorax, or healthy controls (HC). Using 3D Slicer®, we extracted lobar morphometric indices [volume, surface area (SA), surface area-to-volume ratio (SA/V), and shape descriptors] and Hounsfield unit (HU)-derived attenuation metrics from CT datasets. Group differences were assessed per lobe and by within-subject upper-lower asymmetry (Δ analysis). Five side-specific machine learning algorithms were trained to discriminate PT from HC using cross-validated predictions. Youden-optimal receiver operating characteristic (ROC) thresholds were applied to the PE cohort without retraining to quantify pneumothorax-like structural susceptibility.ResultsA total of 561 (right 282 and left 279) side-specific lobes were analyzed. PT showed higher SA/V and less compact lobar geometry than HC and broader high-attenuation tails. In Δ analyses, attenuation-based deltas consistently separated groups, whereas PE showed more geometry-driven asymmetry, including reduced Δ-flatness. support vector machine (SVM) discrimination was the most excellent [area under the curve (AUC): right 0.937; left 0.953]. At Youden thresholds (right 0.811; left 0.841), high-risk/PT-like classifications were enriched in PE versus HC (right 30/67 vs. 1/158; left 38/67 vs. 1/158; both P<0.001).ConclusionsQuantitative CT identifies distinct morphometric–attenuation phenotypes in PT and reveals a substantial PT-like signature in PE, warranting prospective validation with longitudinal pneumothorax outcomes. Highlight box Key findings • Patients with spontaneous pneumothorax (PT) showed a reproducible computed tomography (CT) phenotype with broader Hounsfield unit (HU) distributions, higher surface area-to-volume ratio (SA/V), and less compact upper-lobe geometry. • Upper-lower asymmetry differed by group: attenuation-based Δ metrics most consistently separated PT from both healthy controls (HC) and patients with pectus excavatum (PE), whereas PE was characterized mainly by geometry-driven asymmetry. • In machine-learning analyses, support vector machine (SVM) achieved the best PT-HC discrimination, and transfer classification identified a substantial subset of PE lungs as PT-like. What is known and what is new? • PE has been associated with spontaneous pneumothorax, but the structural basis remains unclear. Prior CT studies have focused mainly on visible blebs/bullae or global density measures. • This study provides quantitative lobe-level CT profiles across PT, PE, and HC using automated lobar segmentation. PT was distinguished mainly by attenuation heterogeneity and cranio-caudal attenuation gradients, whereas PE was distinguished mainly by geometry-related asymmetry. A substantial subset of PE showed PT-like imaging signatures on transfer classification. What is the implication, and what should change now? • Quantitative lobar CT features may help identify a pneumothorax-related structural phenotype beyond conventional visual assessment of blebs/bullae. PE may harbor subclinical PT-like structural vulnerability, but clinical interpretation should remain cautious until prospective multicenter validation confirms prognostic value. Introduction Primary spontaneous pneumothorax (PSP) is relatively common in young adults, particularly in tall, lean males without overt lung disease (1). Its pathogenesis has traditionally been attributed to apical subpleural blebs or bullae; however, accumulating computed tomography (CT) and histopathologic evidence suggests that more subtle, diffuse alterations in lung parenchyma and thoracic architecture may lower the mechanical threshold for pleural rupture (2-4). Despite this, systematic lobe-by-lobe characterization of lung shape and attenuation patterns in this population remains limited. Pectus excavatum (PE), the most common congenital chest wall deformity, has also been associated with an increased risk of spontaneous pneumothorax (2,5-7). Although clinical series report more frequent pneumothorax in patients with pectus deformity, the underlying anatomical mechanism is uncertain (8-10). Specifically, it remains unclear whether anterior chest wall depression secondarily reshapes the lungs—through altered lobar geometry and surface–volume relationships and/or shifts in attenuation distributions—in a manner that resembles the structural substrate observed in PSP (2,5,8). Chest CT provides an appropriate framework to address this gap because it enables not only qualitative assessment of blebs or bullae but also quantitative, lobe-specific characterization of lung morphology and attenuation distributions using Hounsfield unit (HU)-based metrics (11-13). However, few studies have directly compared these CT-derived features across three clinically relevant groups: patients with spontaneous pneumothorax (PT), patients with PE, and healthy controls (HC) with normal lung anatomy (8,14-17). Accordingly, this study had three aims: (I) to compare lobar CT-derived morphometric and HU-based features across groups to identify objective, quantifiable differences in lobe shape and density distributions; (II) to develop a machine-learning model to predict pneumothorax; and (III) to test whether PE exhibits geometric and/or attenuation patterns that shift toward a pneumothorax-like phenotype (18). By doing so, we sought to generate a structural framework that may help explain the reported association between PE and pneumothorax and to identify quantitative imaging features that could support future risk stratification, including machine-learning-based approaches (8). We present this article in accordance with the STROBE reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2926-0589/rc). Methods Study design and subjects At Uijeongbu St. Mary’s Hospital, we conducted a retrospective case-control study from January 2015 to April 2025 involving three parallel cohorts of young adults aged 20–30 years, an age range in which lung growth is generally considered complete (19): (I) PT: patients with a first episode of PSP; (II) PE: patients with PE and no prior history of pneumothorax; and (III) HC: healthy individuals who underwent screening chest CT (primarily as part of routine pre-employment or comprehensive health check-up programs) and showed no clinical or radiological evidence of thoracic pathology. No separate concomitant PT + PE subgroup was included in the present study. For all participants, demographic and clinical covariates were collected, including age, sex, height, weight, and body mass index (BMI). All subjects underwent chest CT using standardized institutional protocols. Exclusion criteria included a history of pulmonary tuberculosis or other chronic lung disease, prior lung, thoracic, and chest wall procedures, and evidence of active pulmonary infection. CT acquisition All chest CT examinations were performed in the supine position at full inspiration. Scans were acquired at 120 kVp with automated tube-current modulation, using a 512×512 matrix and a slice thickness of 1.0 mm, and were reconstructed with a lung kernel. For PT patients, CT scans were typically acquired a few days post-event to allow for clinical stabilization and were performed after near-complete re-expansion (residual collapse ≤5%) and in the absence of atelectasis, with or without an indwelling chest tube, to ensure that lobar morphology reflected baseline lung architecture rather than acute collapse. All Digital Imaging and Communications in Medicine (DICOM) files were anonymized and stored in a secure, access-controlled archive. Image analysis Image analyses were performed using 3D Slicer® (version 5.8.1; r33241), an open-source and extensible platform for medical image informatics, image processing, and 3D visualization, with the Chest Imaging Platform (Lung CT Segmenter module) (11,20,21). DICOM datasets were imported and processed with a fissure-guided algorithm to automatically generate whole-lung and lobar segmentations (11,12,21). All segmentations and primary measurements were obtained fully automatically, without manual editing. From each segmented lobe, morphometric features were automatically extracted (11,22). Statistical analysis All analyses were performed at the side/lobe level because each subject could contribute right- and/or left-sided observations. Continuous variables were summarized as mean ± standard deviation (SD), and categorical variables were summarized as counts and percentages. Normality was assessed using the Shapiro-Wilk test. The primary factor of interest was the study group (HC, PT, and PE), and the outcomes were CT-derived morphometric parameters and attenuation-based features. Primary between-group comparisons were conducted using lobe-level analysis of covariance (ANCOVA), with study group as a fixed effect and age, sex, height, weight, and BMI as covariates. Adjusted, or estimated, marginal means ± standard error (SE) were calculated for each group. When the omnibus group effect was significant (P<0.05), prespecified pairwise comparisons were performed for HC vs. PT, HC vs. PE, and PT vs. PE. Because lobe-wise ANCOVA compares different individuals who may differ in baseline thoracic anatomy, we prespecified a second, within-subject analysis of upper-lower asymmetry to evaluate apical-predominant remodeling. For each hemithorax, Δ was defined as the upper-lobe value minus the lower-lobe value for each key feature. These Δ outcomes were analyzed using the same covariate-adjusted ANCOVA framework and the same multiplicity-control approach as the per-lobe analyses. This Δ approach reduces shared subject- and scanner-level variability and enables between-group comparisons of the magnitude and direction of apical-basal gradients among HC, PT, and PE. Machine-learning workflow For risk-oriented analyses, we developed side-specific machine-learning classifiers in Orange3® (Orange Data Mining; v3.38.0), an open-source visual programming environment for data mining and machine learning, to distinguish PT from HC using predefined morphometric and attenuation features, supplemented with clinical covariates when available. Logistic regression, random forest, gradient boosting, support vector machine (SVM), and neural network classifiers were evaluated using 10-fold cross-validation. Model performance was assessed using the area under the curve (AUC), accuracy, precision, recall, and F1 score. The model achieving the highest cross-validated AUC was refit using the full PT + HC dataset and was then applied to the PE cohort without retraining to generate predicted probabilities for each PE case, defined as pneumothorax risk scores. High-risk status was defined using the Youden-optimal cutoff derived from the PT-HC receiver operating characteristic (ROC) curve. Because this threshold was derived internally from the same dataset, the resulting high-risk proportions were interpreted as indicators of predicted structural vulnerability, or PT-likeness, rather than direct estimates of clinical incidence. Orange3® implementation Right- and left-lung feature tables were imported into Orange3®, harmonized by standardizing feature labels, and concatenated into a single dataset. Classifiers were trained to distinguish PT from HC. For PT, only the ipsilateral (affected) side was included, whereas both sides were included for HC and PE. In a secondary subject-level analysis, each participant’s risk score was defined as the higher of the two side-specific predicted probabilities, and risk-score distributions were compared across HC, PE, and PT. Ethics The study protocol was approved by Institutional Review Board of Uijeongbu St. Mary’s Hospital, The Catholic University of Korea (No. UC26RISI0015). The requirement for informed consent was waived for the de-identified control cohort. All imaging and clinical data were de-identified before analysis, and analysis scripts and segmentation masks will be made available upon publication under an institutional data-use agreement. The study was conducted in accordance with the principles of the Declaration of Helsinki and its subsequent amendments. Results When stratified by side, the right hemithorax consisted of 158 HC (132 males, 26 females; mean age 26.96±2.36 years), 57 PT (54 males, 3 females; 24.58±3.44 years), and 67 PE (40 males, 27 females; 23.28±3.50 years), and the left hemithorax likewise included 158 HC (132 males, 26 females; mean age 26.96±2.36 years), 54 PT (49 males, 5 females; mean age 23.89±3.23 years), and 67 PE (40 males, 27 females; 24.58±3.44 years). All subsequent analyses were conducted at the side/lobe level. The CT-derived outcomes comprised morphometric parameters—volume (cm3), surface area (SA) (mm2), elongation, flatness, and surface area-to-volume ratio (SA/V)—and attenuation features—mean, 5th percentile (HU5), 95th percentile (HU95), range, and Bowley skewness (24). Analyses of morphometric parameters and attenuation features in the groups Detailed lobar morphometric parameters and attenuation features for the HC, PT, and PE groups are presented in Table 1. These descriptive data include lobar volume, SA, roundness, elongation, flatness, SA/V, mean attenuation, HU5, HU95, HU range, and Bowley skewness for the right and left upper and lower lobes. Table 1Morphometric parameters and attenuation featuresParametersRULRLLLULLLLHC Volume (cm3)873.7±179.81,247.3±300.51,116.7±232.31,181.0±307.0 SA (mm2)60,252.6±8,258.884,295.0±13,445.676,549.9±10,580.482,083.3±13,890.4 Roundness0.8±0.00.8±0.00.7±0.01.5±0.2 Elongation1.3±0.11.3±0.21.9±0.10.8±0.0 Flatness1.5±0.11.2±0.11.5±0.11.4±0.1 SA/V70.1±6.669.1±6.769.8±6.671.3±7.4 Mean HU−809.9±31.4−792.8±47.8−814.6±33.2−788.3±49.3 HU5−894.5±23.7−887.9±32.4−902.1±24.3−887.3±33.9 HU95−579.1±48.3−525.7±73.9−560.3±52.4−502.2±71.9 HU range315.4±30.9362.2±45.2341.8±33.3385.1±41.9 Skewness0.7±0.10.7±0.10.7±0.00.7±0.1PT Volume (cm3)814.0±230.11,075.1±327.71,072.2±249.7994.4±284.4 SA (mm2)64,862.0±20,056.183,854.1±24,472.577,166.6±12,855.375,085.6±13,104.8 Roundness0.7±0.10.7±0.10.7±0.10.7±0.1 Elongation1.4±0.21.5±0.22.0±0.31.5±0.3 Flatness1.5±0.21.2±0.11.5±0.21.4±0.2 SA/V81.4±16.380.1±14.973.3±9.578.4±13.8 Mean HU−798.1±42.5−755.4±59.7−784.6±38.4−732.6±70.5 HU5−912.5±58.7−906.3±63.7−906.1±57.3−877.1±65.1 HU95−530.0±96.8−371.7±197.6−481.4±78.4−376.4±140.5 HU range382.6±118.6534.6±220.0424.7±98.7500.7±137.5 Skewness0.6±0.20.6±0.10.6±0.10.6±0.1PE Volume (cm3)673.1±144.21,062.4±210.8866.3±202.0820.8±242.9 SA (mm2)51,855.9±8,534.577,537.3±11,343.865,291.5±9,249.967,607.5±12,566.4 Roundness0.8±0.10.8±0.10.8±0.10.7±0.1 Elongation1.3±0.11.4±0.22.1±0.21.7±0.2 Flatness1.4±0.21.3±0.11.4±0.11.6±0.2 SA/V78.4±11.274.0±6.677.6±12.285.9±15.8 Mean HU−806.8±22.6−799.9±29.8−803.9±27.9−772.9±40.7 HU5−907.3±23.0−911.1±24.1−911.5±28.5−900.0±28.4 HU95−565.6±48.4−516.6±76.1−535.4±55.9−458.5±82.2 HU range341.7±55.2394.5±79.7376.1±59.0441.5±78.2 Skewness0.6±0.10.7±0.10.6±0.10.6±0.1Data are presented as mean ± standard deviation. HC, healthy controls; HU, Hounsfield unit; HU5, 5th percentile; HU95, 95th percentile; LLL, left lower lobe; LUL, left upper lobe; PE, patients with pectus excavatum; PT, patients with spontaneous pneumothorax; RLL, right lower lobe; RUL, right upper lobe; SA, surface area; SA/V, surface area-to-volume ratio. Morphometric parameters and attenuation features Data are presented as mean ± standard deviation. HC, healthy controls; HU, Hounsfield unit; HU5, 5th percentile; HU95, 95th percentile; LLL, left lower lobe; LUL, left upper lobe; PE, patients with pectus excavatum; PT, patients with spontaneous pneumothorax; RLL, right lower lobe; RUL, right upper lobe; SA, surface area; SA/V, surface area-to-volume ratio. Comparisons among HC, PT, and PE Per-lobe comparisons After covariate adjustment, significant between-group differences were observed across multiple morphometric and attenuation features in all lobes (Table 2). Across lobes, PT was generally characterized by less compact geometry, higher SA/V, and broader HU distributions than HC. PE also differed from HC, but its pattern was more heterogeneous and was driven more by shape-related features than by the consistent attenuation shifts seen in PT. Overall, SA/V, HU95, and HU range provided some of the most reproducible separation between PT and HC, whereas PE often showed intermediate values or lobe-specific deviations. Table 2Per-lobe between-group comparisons of CT-derived morphometric and attenuation featuresLobeVariablesEstimated marginal meansP valueP value (post-hoc test)HCPTPEHC vs. PTHC vs. PEPT vs. PERULRoundness0.760.690.74<0.001****Flatness1.511.511.430.02**Elongation1.321.371.320.005**Volume850.39768.89751.090.004**Mean HU−812.40−795.73−802.810.01**HU5−894.65−909.52−909.510.02*HU95−587.45−528.55−547.14<0.001****HU range307.21380.97368.37<0.001****SA58,643.8663,757.9656,588.940.002***SA/V69.7782.9877.85<0.001*****Skewness0.680.610.62<0.001****RLLRoundness0.760.690.72<0.001***Flatness1.201.201.27<0.001****Elongation1.351.481.41<0.001**Volume1,222.601,020.201,167.3500<0.001**Mean HU−799.96−747.40−789.81<0.001***HU5−889.84−901.09−910.970.02*HU95−543.82−355.51−487.61<0.001***HU range376.46491.92457.44<0.001****SA82,429.6581,752.2683,724.310.80SA/V66.4882.2973.44<0.001******Skewness0.7000.650.64<0.001**LULRoundness0.740.730.740.86Flatness1.511.541.43<0.001****Elongation1.951.972.010.22Volume1,078.951,034.49985.600.04*Mean HU−816.93−778.91−802.96<0.001***HU5−905.19−899.84−909.260.37HU95−564.86−473.06−531.49<0.001******HU range340.33426.78377.77<0.001******SA74,865.7174,903.0471,087.460.04SA/V70.5374.1075.160.005***Skewness0.700.650.650.001****LLLRoundness0.750.730.720.01Flatness1.391.561.55<0.001****Elongation1.501.621.64<0.001****Volume1,147.73955.45929.37<0.001****Mean HU−789.26−727.83−774.32<0.001****HU5−889.62−870.46−899.810.001***HU95−502.39−377.23−461.45<0.001****HU range387.23498.23438.36<0.001******SA80,397.0272,670.6673,494.620.001****SA/V71.8779.0883.83<0.001****Skewness0.710.650.62<0.001*****, P<0.05; **, P<0.01. CT, computed tomography; HC, healthy controls; HU, Hounsfield unit; HU5, 5th percentile; HU95, 95th percentile; LLL, left lower lobe; LUL, left upper lobe; PE, patients with pectus excavatum; PT, patients with spontaneous pneumothorax; RLL, right lower lobe; RUL, right upper lobe; SA, surface area; SA/V, surface area-to-volume ratio. Per-lobe between-group comparisons of CT-derived morphometric and attenuation features *, P<0.05; **, P<0.01. CT, computed tomography; HC, healthy controls; HU, Hounsfield unit; HU5, 5th percentile; HU95, 95th percentile; LLL, left lower lobe; LUL, left upper lobe; PE, patients with pectus excavatum; PT, patients with spontaneous pneumothorax; RLL, right lower lobe; RUL, right upper lobe; SA, surface area; SA/V, surface area-to-volume ratio. Within-subject upper-lower asymmetry (Δ upper-lower, covariate-adjusted) Results of the within-subject Δ analyses are summarized in Table 3. On both sides, PT showed the most consistent attenuation-based asymmetry, with significant differences in Δ-mean HU, Δ-HU95, and Δ-HU range compared with HC and, in several cases, PE. In contrast, PE was characterized more prominently by geometry-related asymmetry, particularly in Δ-flatness and related shape indices. These findings suggest that attenuation-gradient heterogeneity is a stronger distinguishing feature of PT, whereas PE is more strongly associated with geometry-driven upper-lower asymmetry. Table 3Within-subject upper-lower asymmetry in CT-derived featuresSideVariablesEstimated marginal meansP valueP value (post-hoc test)HCPTPEHC vs. PTHC vs. PEPT vs. PERightRoundness−0.003−0.0010.0120.21Flatness0.2990.3130.169<0.001****Elongation−0.035−0.108−0.0960.04Volume−372.21−233.31−416.25<0.001****Mean HU−12.44−48.33−13.00<0.001****HU5−4.812−8.4381.4660.002**HU95−43.621−173.042−59.530<0.001****HU range−38.809−164.604−60.997<0.001****SA−23,785.80−17,994.30−27,135.37<0.001***SA/V1.2970.6824.4100.057Skewness−0.019−0.037−0.0220.28LeftRoundness−0.0160.0000.0160.02*Flatness0.1230.182−0.124<0.001**Elongation0.4560.3500.3740.02*Volume−70.33864.86954.199<0.001****Mean HU−27.631−51.003−28.604<0.001****HU5−15.402−27.998−9.233<0.001****HU95−62.797−104.744−70.517<0.001****HU range−47.395−76.748−61.2840.010**SA−5,609.4551,637.239−2,506.440<0.001**SA/V−1.320−4.794−8.642<0.001****Skewness−0.008−0.0010.0230.08*, P<0.05; **, P<0.01. CT, computed tomography; HC, healthy controls; HU, Hounsfield unit; HU5, 5th percentile; HU95, 95th percentile; PE, patients with pectus excavatum; PT, patients with spontaneous pneumothorax; SA, surface area; SA/V, surface area-to-volume ratio. Within-subject upper-lower asymmetry in CT-derived features *, P<0.05; **, P<0.01. CT, computed tomography; HC, healthy controls; HU, Hounsfield unit; HU5, 5th percentile; HU95, 95th percentile; PE, patients with pectus excavatum; PT, patients with spontaneous pneumothorax; SA, surface area; SA/V, surface area-to-volume ratio. Morphological and attenuation signatures underlying pneumothorax susceptibility In covariate-adjusted per-lobe comparisons, PT showed widespread differences from HC across both morphometric and attenuation features in multiple lobes (3,25). Morphometrically, PT was characterized by less compact lobar geometry, with lower roundness, higher elongation, and consistently higher SA/V; among these, SA/V provided the most consistent separation from HC across lobes. In contrast, differences in lobar volume and SA were lobe-specific rather than uniform throughout the lung. Attenuation features also differed across several lobes, with PT showing higher HU95 values and wider HU ranges, consistent with broader HU distributions. In the within-subject Δ(upper-lower) analysis, attenuation-based delta metrics—Δmean HU, ΔHU95, ΔHU range, and, where applicable, ΔHU5—showed the most consistent between-group differences, with PT differing from both HC and PE. PE showed a distinct pattern. In per-lobe analyses, PE differed from HC in selected morphometric features, particularly shape descriptors and SA/V, with these differences being more prominent in the lower lobes, where SA/V tended to be higher (26). In the Δ analysis, PE showed strong group effects for geometry-based asymmetry measures, including Δ-flatness, as well as side-dependent patterns in ΔSA/V (25). By contrast, attenuation-based Δ metrics in PE were generally closer to HC than to PT. Overall, both PT and PE differed from HC in absolute lobar measurements and in upper-lower asymmetry. However, PT was more consistently distinguished by attenuation-based features and attenuation-gradient deltas, whereas PE was more consistently distinguished by geometry-related asymmetry metrics. Prediction of pneumothorax with machine-learning algorithms using morphometric and attenuation features Because the predefined morphometric and attenuation features showed reproducible PT-HC separation across multiple lobes and in upper-lower (Δ) analyses, we used these variables to develop side-specific machine-learning classifiers. Clinical covariates (age, sex, height, weight, and BMI) were also included in all models. Models were implemented in Orange® and evaluated using five algorithms (23). For right-sided pneumothorax, the SVM achieved the best performance (AUC =0.937; accuracy =0.902; F1-score =0.899; precision =0.902; recall =0.902). For left-sided pneumothorax, SVM likewise performed best (AUC =0.953; accuracy =0.915; F1-score =0.911; precision =0.916; recall =0.915) (Table 4). ROC curves are shown in Figure 1. Table 4Performance of machine-learning classifiers for PT vs. HC discrimination using morphometric and attenuation CT featuresSideModelAUCAccuracyF1 scorePrecisionRecallRightSVM0.9370.9020.8990.9020.902Neural network0.9370.8880.8860.8860.888Random forest0.9240.8740.8700.8710.874Logistic regression0.9120.8980.8960.8960.898Gradient boosting0.8850.8880.8840.8870.888LeftSVM0.9530.9150.9110.9160.915Gradient boosting0.9480.9010.8990.8990.901Neural network0.9440.9150.9130.9140.915Random forest0.9170.8920.8860.8900.892Logistic regression0.9080.9010.9000.9000.901AUC, area under the curve; CT, computed tomography; HC, healthy controls; PT, patients with spontaneous pneumothorax; SVM, support vector machine. Performance of machine-learning classifiers for PT vs. HC discrimination using morphometric and attenuation CT features AUC, area under the curve; CT, computed tomography; HC, healthy controls; PT, patients with spontaneous pneumothorax; SVM, support vector machine. Receiver operating characteristic curves for machine-learning prediction of pneumothorax: (A) right side and (B) left side. FP, false positive; SVM, support vector machine; TP, true positive. Transfer classification of PE lungs using a pneumothorax-trained model When the side-specific PT-HC-trained SVMs were applied to the PE cohort without retraining, a substantial proportion of PE lungs were classified as PT-like high-risk at the Youden-optimal thresholds (right 0.811; left 0.841). Specifically, 30 of 67 PE patients [44.78%; 95% confidence interval (CI): 33.48–56.64%] met the PT-like high-risk criterion on the right side and 38 of 67 (56.72%; 95% CI: 44.81–67.90%) on the left side. Compared with HC, PE showed markedly higher rates of PT-like high-risk classification, with relative risks of 70.75 (95% CI: 9.85–508.20) for the right and 89.61 (95% CI: 12.56–639.32) for the left (both P<0.001, Table 5). Table 5PT-like classifications in PE using a PT-HC-trained side-specific modelSideYouden thresholdPT-like high-risk, n/N (%)PE (95% CI), %Relative risk (95% CI)P valueRight0.81130/67 (44.78)33.48–56.6470.75 (9.85–508.20)<0.001Left0.84138/67 (56.72)44.81–67.9089.61 (12.56–639.32)<0.001CI, confidence interval; HC, healthy controls; PE, pectus excavatum; PT, patients with spontaneous pneumothorax. PT-like classifications in PE using a PT-HC-trained side-specific model CI, confidence interval; HC, healthy controls; PE, pectus excavatum; PT, patients with spontaneous pneumothorax. Discussion In this retrospective study of young adults, we compared lobe-wise CT-derived morphometric and attenuation characteristics among PT, PE, and HC, and then applied machine-learning classifiers trained on PT versus HC to assess whether PE exhibits pneumothorax-like imaging signatures that may help explain its increased susceptibility to pneumothorax (10,27). Across lobes, PT demonstrated the most pervasive deviation from HC, spanning both morphology and attenuation (2). Morphometrically, PT consistently exhibited a less compact lobar configuration accompanied by relative surface enrichment, with SA/V providing the clearest and most reproducible separation from HC across lobes (4,9). Concomitant differences in absolute volume and SA were lobe-dependent, arguing against a uniform scaling effect and instead supporting heterogeneous remodeling across the lung. In parallel, PT showed a uniform attenuation phenotype characterized by broader HU distributions, including higher HU95 and a wider HU range across multiple lobes, consistent with increased parenchymal heterogeneity rather than a single focal abnormality. Importantly, the within-subject Δ(upper-lower) framework—designed to reduce between-subject baseline variability—highlighted cranio-caudal attenuation gradients as a core feature of PT: attenuation-based deltas most consistently separated PT from both HC and PE, indicating a distinctive upper-lower density gradient beyond absolute lobar differences. PE, however, did not simply mirror HC. In per-lobe comparisons, PE showed selective differences in shape descriptors and SA/V, with notable deviations in lobes where thoracic geometry may constrain lung expansion (18,28). The Δ analyses further clarified that PE’s dominant signal was geometry-driven asymmetry: PE showed strong group effects in Δ-flatness on both sides and side-dependent ΔSA/V tendencies, consistent with altered surface-size relationships and lobar shape expression related to chest wall constraint (28). In contrast, attenuation-gradient measures in PE generally tracked closer to HC than did those in PT, supporting the interpretation that PE and PT diverge from the control pattern along partly different phenotypic axes—PE predominantly through geometric remodeling and PT through a stronger attenuation-gradient (density heterogeneity) signature. Rather than simply indicating that lung geometry differs in PE, our findings support the hypothesis that anterior chest wall depression produces a regionally constrained lobar configuration—reflected by altered SA/V, shape indices, and upper-lower asymmetry—that may redistribute subpleural mechanical stress and thereby create a permissive structural background for pneumothorax. The machine-learning analyses further supported this multivariable distinction (23). Side-specific SVMs trained on PT versus HC showed excellent discrimination in internal validation. When these models were applied to the PE cohort, a substantial proportion of PE cases were classified as PT-like high-risk. Relative to HC, PE exhibited markedly higher PT-like classification rates, suggesting that within the PT-HC-defined feature space, many PE lungs cluster closer to the PT-like phenotype than to the control-like phenotype (5). Importantly, because these classifiers were trained on cross-sectional imaging signatures rather than longitudinal outcomes, the resulting probabilities should be interpreted as imaging-based PT-likeness (predicted structural vulnerability) rather than estimates of future clinical incidence. From a clinical standpoint, these findings suggest that PT and PE are associated with partly distinct quantitative CT phenotypes, with PT showing stronger attenuation-gradient heterogeneity and PE showing more prominent geometry-related asymmetry. While altered lung geometry in PE is a widely recognized phenomenon, our quantitative analysis provides a specific mechanical framework for its association with pneumothorax. We hypothesize that the chest wall deformity in PE does not merely alter the gross shape of the lungs, but specifically forces the lobar architecture into higher SA/V configurations with pronounced cranio-caudal asymmetry (e.g., altered flatness gradients). This specific structural shift intrinsically amplifies baseline mechanical stress and distending forces across the subpleural parenchyma. By lowering the baseline mechanical threshold for pleural rupture, this PT-like geometric remodeling creates a highly susceptible environment where secondary events—such as abrupt intrathoracic pressure changes, strenuous activity, or focal subpleural fragility—can more readily precipitate overt pleural disruption compared to anatomically normal chests. This proposed pathway is consistent with our observation that PE shows geometry-linked asymmetry while PT exhibits a more pronounced attenuation-gradient heterogeneity signature, suggesting partially distinct but potentially converging mechanisms of susceptibility. If confirmed in external cohorts, quantitative CT features—such as lobar SA/V, shape indices, and upper–lower asymmetry metrics—could support imaging-based phenotyping and hypothesis-driven risk stratification in PE. For example, PE patients exhibiting a strongly PT-like signature may warrant closer follow-up and counseling regarding potential precipitating factors, whereas those with a more control-like pattern may require less intensive surveillance. At present, however, these implications remain hypothesis-generating and should await prospective validation with incident pneumothorax outcomes. Several limitations should be acknowledged. First, this was a single-center retrospective study of young adults, and selection bias cannot be excluded. Although CT acquisition and reconstruction were relatively homogeneous, they were not completely uniform, and residual technical variability may have influenced absolute attenuation values. Second, PT patients were scanned after chest tube drainage and near-complete re-expansion, so the observed PT phenotype may partly reflect post-event or treatment-related changes rather than purely pre-collapse architecture. We did not separately adjust for the presence of an indwelling chest tube or the residual degree of lung collapse, both of which may have affected HU-based attenuation metrics and SA/Vs by altering local parenchymal expansion and regional geometry. Therefore, the more pronounced attenuation gradient heterogeneity observed in PT should be interpreted with caution. Although CT acquisition was intentionally delayed by several days to allow partial stabilization, scans were still obtained during the acute or subacute phase, often with an indwelling chest tube in place. Subtle post-collapse changes, including micro-atelectatic effects or localized parenchymal responses related to the tube, may have accentuated the observed heterogeneity and may not fully reflect baseline anatomical variation alone. Third, lobar segmentation and feature extraction were fully automated without manual correction of every lobe, and minor segmentation errors may have affected surface- and shape-based metrics. The machine-learning analysis also has important limitations. All models were internally validated only, and no external validation was performed; accordingly, the reported performance metrics should be interpreted as exploratory and hypothesis-generating. Because the side-specific SVM models were applied to the PE cohort without external recalibration, overfitting cannot be excluded. More importantly, longitudinal follow-up was unavailable, so it was not possible to determine which PE patients subsequently developed pneumothorax, thereby precluding individual-level outcome prediction. In addition, a marked sex imbalance was present across the cohorts. Although sex was included as a covariate in the statistical analyses and as a clinical feature in the machine-learning models, residual confounding due to sex-related differences in thoracic and pulmonary anatomy cannot be excluded. Because sex-stratified or sex-matched sensitivity analyses were not performed, part of the observed discriminative performance may reflect sex-associated morphological differences rather than PT-related imaging phenotype alone. A clinically calibrated prognostic score for future pneumothorax occurrence would be a valuable extension of this work. However, although the present model yields imaging-based PT-like probabilities, these outputs reflect structural vulnerability rather than true incident risk because longitudinal outcome data were unavailable. A logical next step will be to develop and externally validate a prospective quantitative risk model integrating CT-derived features, demographic variables, and incident pneumothorax outcomes. Conclusions In young adults, both PT and PE exhibited quantitative CT phenotypes distinct from that of HC. PT was characterized more strongly by attenuation heterogeneity and cranio-caudal density gradients, whereas PE was characterized more consistently by geometry-related asymmetry. The PT-like imaging phenotype observed in a subset of the PE cohort may help inform future hypothesis-driven research. However, larger multicenter studies with harmonized imaging protocols and longitudinal outcome follow-up are required before these imaging signatures can be interpreted in prognostic or mechanistic terms or used to inform clinical decision-making. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2926-0589/prf