The Diagnostic Dilemma Bronchiolar adenoma (BA) is a rare benign pulmonary tumor that looks remarkably similar to lung adenocarcinoma (LAC) on CT imaging and even on intraoperative frozen section microscopy. This mimicry leads to unnecessary escalation of surgery - patients get lobectomies when wedge resections would suffice.
Why BA Is Hard to Diagnose The defining histological feature of BA is a double-layered cellular structure (basal cells plus luminal cells). However, detecting this requires immunohistochemical staining with markers like p40 and CK5/6, which are not available during intraoperative frozen section analysis. Surgeons must make decisions without definitive diagnosis.
The AI Histogram Approach This study leveraged a commercial AI lung nodule platform (Deepwise Technology) that automatically segments nodules on CT and extracts 17 quantitative histogram parameters - from simple size measurements to texture features like entropy, kurtosis, and CT value variance. These AI-derived features were used to build a nomogram distinguishing BA from LAC.
Two-Center Design To ensure generalizability, 215 patients from two independent hospitals were enrolled - 151 from the first center as the training cohort and 64 from the second as an independent external validation cohort. This is far larger than the previous only study on CT-based BA diagnosis (15 patients from a single center).
AI Platform and Feature Extraction All CT images (DICOM format) were processed by the Deepwise AI lung nodule platform, which uses a recurrent CNN to automatically segment each nodule and calculate 17 histogram parameters for the entire nodule volume. Parameters included 2D and 3D size measurements, CT value statistics (mean, variance, max, min), and texture descriptors (entropy, kurtosis, skewness, sphericity, compactness, energy).
Feature Selection Process Univariate logistic regression identified 12 initial predictors. Backward stepwise multivariate logistic regression then reduced these to 3 independent predictors: nodule density (solid vs. ground-glass vs. part-solid), 2D short diameter, and CT value variance. These three features formed the basis of the nomogram.
Statistical Validation Model performance was evaluated by 5-fold cross-validation in the training cohort, then independently assessed in the external validation cohort from the second hospital. Calibration curves, ROC curves, and decision curve analysis were used to assess accuracy, calibration, and clinical utility.
Nomogram Formula The final predictive model was: Logit(P) = -1.120 + 0.795 x Density + 0.231 x 2D short diameter - 0.00002 x CT value variance. This simple three-variable equation can be quickly calculated by any clinician from AI-extracted CT parameters.
Size Differences Are Significant LAC nodules were significantly larger than BA nodules across all size metrics. Median 2D short diameter was 6.0 mm for BA versus 9.0 mm for LAC in the training cohort (p less than 0.001), and 3D volume was 180 mm3 for BA versus 888 mm3 for LAC (p less than 0.001). BA tends to be smaller and more compact.
CT Value Variance: The Key Discriminator CT value variance - measuring how much CT attenuation values vary within the nodule - was significantly higher in BA than LAC (median 65,120 vs. 43,044 HU2, p = 0.001). This counter-intuitive finding suggests BA has greater internal heterogeneity in tissue density, likely reflecting its double-layered cellular structure with both mucin-filled and solid regions.
Nodule Density Pattern Pure ground-glass and part-solid nodules were more commonly associated with LAC, while solid density was more often seen in BA. Density was a significant predictor, reflecting different underlying pathological mechanisms.
Entropy Is Higher in LAC Entropy, which measures the complexity and randomness of image texture, was significantly higher in LAC (median 8.90 vs. 7.84 in training cohort, p less than 0.001). Higher entropy reflects greater textural heterogeneity, consistent with the infiltrative and architecturally disordered nature of adenocarcinoma.
Training Cohort Performance The three-variable nomogram achieved an AUC of 0.821 (95% CI: 0.753-0.890) in the training cohort, with 5-fold cross-validation yielding AUC of 0.81 +/- 0.06 - indicating the model is stable and not overfitting.
External Validation Performance In the independent second-center validation cohort, the AUC was 0.811 (95% CI: 0.693-0.928). The minimal drop from training to validation AUC confirms generalizability across different hospitals, scanner brands, and patient populations.
Calibration Is Good Calibration curve analysis showed good agreement between predicted probabilities and observed outcomes in both cohorts (Hosmer-Lemeshow p = 0.100 for training, 0.200 for validation), confirming that the model's risk scores are clinically interpretable.
Clinical Utility by Decision Curve Analysis Decision curve analysis demonstrated positive net benefit within a clinically relevant threshold range, confirming the nomogram adds value over default treat-all or treat-none strategies when guiding surgical approach decisions.
Retrospective Design and Sample Size This was a retrospective study limited to 215 patients total. BA is rare, and the sample size restricts the robustness of the model, particularly for rarer BA subtypes. Larger prospective multicenter studies are needed.
Single Nodule per Patient Only the largest nodule per patient was analyzed. In patients with multiple nodules, smaller synchronous nodules (which may have different characteristics) were excluded, potentially limiting generalizability.
Commercial AI Platform Dependency The histogram features were automatically extracted using a specific commercial AI platform (Deepwise). Replicating results requires access to this or a similar tool, limiting reproducibility in centers without this technology.
Clinical Impact If validated prospectively, this tool could directly change surgical decision-making. Surgeons who receive a high BA probability score preoperatively could plan for limited wedge resection, avoiding unnecessary extended surgery, reducing complications, shortening hospital stays, and improving patient quality of life.