Clinical Challenge: Micropapillary (MP) and solid (SP) patterns in lung adenocarcinoma are aggressive histological subtypes associated with recurrence, lymph node metastasis, and poor survival, but are currently only confirmed by postoperative pathology.
Study Goal: This retrospective study developed and validated a clinical-radiomics nomogram to preoperatively identify the presence of high-risk MP/SP components in LUAD, enabling more personalized surgical and adjuvant therapy planning.
Study Design: 180 surgically confirmed NSCLC patients (Stages I-IIIA) were enrolled, with 70% (n=126) forming the training cohort and 30% (n=54) the validation cohort.
Best Model Performance: The comprehensive clinical-radiomics model achieved the highest AUC of 0.9186 (training) and 0.9396 (validation), significantly outperforming the clinical-only (AUC 0.7975/0.8462) and radiomics-only (AUC 0.8896/0.8901) models.
WHO Classification: The 2021 WHO classification of lung tumors recommends quantifying each histological component of LUAD in 5% increments, with the predominant pattern guiding prognosis and management decisions.
MP/SP Aggressiveness: A meta-analysis of 19,502 LUAD patients across 48 studies confirmed that MP and SP components correlate with increased recurrence, lymph node metastasis, and significantly reduced overall survival.
Treatment Implications: Identifying MP/SP pre-operatively is critical because these high-risk patterns influence the decision between sublobar resection and lobectomy, and determine the need for adjuvant chemotherapy.
Current Limitation: Histopathologic evaluation requires invasive procedures (biopsy or resection) and may not capture tumor heterogeneity; preoperative biopsy specimens often miss areas of aggressive MP/SP histology present only in portions of the tumor.
CT Protocol: All CT scans were acquired on a 640-slice multidetector scanner (Canon Aquilion ONE) with standardized parameters including 120 kVp, 1mm isotropic voxels, FC81 reconstruction kernel, and median CTDIvol 3.2 mGy.
Radiomics Workflow: IBSI-compliant feature extraction using PyRadiomics on 3D Slicer-segmented ROIs yielded features including first-order statistics, 2D/3D shape descriptors, and texture features from GLCM, GLRLM, GLSZM, and NGTDM matrices, plus wavelet transformations.
Feature Selection: Z-score normalization followed by LASSO regression with 10-fold cross-validation selected the most predictive radiomic features. ICC > 0.8 was required for intra- and inter-observer reproducibility to retain features.
Three Models Compared: A clinical model (based on multivariate logistic regression of significant CT morphological features), a radiomics model (RadScore from LASSO-selected features), and a comprehensive ComScore model integrating both.
Significant Clinical Features: Nodule size (p = 0.010), lobulation (p = 0.009), spiculation (p = 0.005), vacuole sign (p = 0.017), pleural indentation (p < 0.01), and vascular abnormality (p = 0.026) were significantly associated with high-risk MP/SP patterns in training.
Consistent Validation: These associations remained significant in the validation cohort for nodule size, lobulation, spiculation, vacuole sign, pleural indentation, and vascular abnormality, confirming their reliability as MP/SP predictors.
Radiomics Features: LASSO-selected radiomics features captured textural heterogeneity and morphological complexity within the tumor that correlate with the disordered growth architecture of MP and SP histological subtypes.
Non-Significant Factors: Age, gender, BMI, nodule type, and smoking history did not reach significance in the training cohort, suggesting that imaging characteristics are more informative than demographic risk factors for subtype prediction.
Comprehensive Model Superiority: The combined ComScore model achieved AUC 0.9186 (training) and 0.9396 (validation), significantly outperforming both the clinical (0.7975/0.8462) and radiomics-only (0.8896/0.8901) models (DeLong test, p < 0.05).
Decision Curve Analysis: DCA demonstrated greater net clinical benefit for the comprehensive model across a wider range of threshold probabilities, confirming that integrating clinical and radiomics information improves real-world decision value.
Calibration: Calibration curves based on 1,000 bootstrap resamples showed good agreement between predicted probabilities and observed outcomes for all three models, indicating reliable probability estimation.
Optimal Threshold: The Youden index-derived threshold provided the best balance of sensitivity and specificity for the nomogram, with corresponding PPV, NPV, accuracy, and F1 score computed for clinical reference.
Treatment Personalization: Preoperative identification of MP/SP-positive cases through the nomogram could trigger more extensive resection planning, lymph node dissection strategy, and multidisciplinary team consultation before patients enter the operating room.
Adjuvant Therapy Guidance: Patients with predicted high-risk MP/SP patterns could be considered for postoperative adjuvant chemotherapy protocols even after complete resection of apparent early-stage disease, reflecting the higher recurrence risk.
Patient Counseling: The nomogram score could be shared with patients during preoperative consultations to provide more accurate prognostic information and facilitate shared decision-making about surgical extent and post-surgical surveillance.
Non-Invasive Advantage: By using only CT imaging and clinical data - already available from standard preoperative workup - the nomogram adds predictive value without requiring additional tests, biopsies, or patient burden.
Single-Center Retrospective Design: All 180 patients came from one institution, raising concerns about CT protocol specificity and the generalizability of the radiomics features to other scanners and reconstruction settings.
Sample Size: With 180 patients total (126 training, 54 validation), statistical power is limited and the model is vulnerable to overfitting, particularly for the radiomics component with its high feature dimensionality.
Manual Segmentation: Manual tumor ROI delineation introduces inter-observer variability that could affect feature reproducibility in multicenter settings. Automated segmentation would be required for practical clinical deployment.
Multicenter External Validation: Independent validation across multiple institutions with diverse patient populations and CT equipment is essential before the nomogram can be recommended for routine clinical preoperative assessment.