Multiple subsolid nodules present a growing clinical challenge. The widespread adoption of low-dose CT for lung cancer screening has led to increasing detection of subsolid nodules, which include pure ground-glass nodules and part-solid nodules. In patients with multiple such nodules, the pattern more likely reflects multicentric lung cancer rather than metastatic disease.
When a dominant tumor is surgically removed, other nodules often cannot be simultaneously resected. These remaining residual nodules require careful surveillance over time. Guidelines recommend regular CT monitoring for at least 5 years, but clinical tools to identify which residual nodules are likely to grow remain limited.
Identifying high-risk residual nodules preoperatively would allow surgeons to plan concurrent resections when feasible, intensify postoperative surveillance for specific patients, and avoid unnecessary interventions for stable nodules. Current morphological and quantitative CT features provide some guidance, but their accuracy is limited and dependent on radiologist experience.
Radiomics offers a data-driven approach that can extract hundreds of quantitative imaging features invisible to the naked eye, potentially capturing tumor heterogeneity signatures that predict future growth. This study is among the first to apply radiomics specifically to predict 5-year growth of residual nodules in this surgical context.
A large surgical cohort with long follow-up. Records from 1,392 patients who underwent resection for lung subsolid nodules confirmed as adenocarcinoma or precursor lesions between 2014 and 2018 were reviewed. After applying inclusion and exclusion criteria, 208 patients with 603 residual nodules and follow-up exceeding 5 years were included.
Residual nodules were defined as subsolid nodules on preoperative CT that persisted after dominant tumor resection, with sizes ranging from 3 to 20 mm and a solid component less than 5 mm. Preoperative CT scans served as baseline, and CT taken at 5 years post-resection served as the reference point for growth assessment.
Growth was defined by any of three criteria: an increase in maximum diameter of 2 mm or more, an increase in solid component size of 2 mm or more, or the appearance of a new solid component. Nodules meeting any of these criteria at or within 5 years of surgery were classified as grown; all others were stable.
Patients harbored 1 to 26 residual nodules each, reflecting the heterogeneity of multiple-nodule burden in this population. The 603 nodules were randomly split into training (498 nodules) and testing (105 nodules) sets at approximately a 4:1 ratio, while maintaining similar growth-to-stable proportions in each set.
2,264 radiomics features extracted per nodule. Using the uAI Research Portal platform with an AI-assisted segmentation workflow, 3D volumes of interest were delineated for each residual nodule. PyRadiomics extracted 2,264 features per nodule from the original image and 24 processed filter images including wavelet, Laplacian, and others.
Feature categories included 14 shape-based features, and texture and first-order statistics from each image filter, covering gray-level co-occurrence matrices, gray-level run-length matrices, gray-level size zone matrices, gray-level dependence matrices, and neighborhood gray-tone difference matrices.
Feature selection began by excluding features with intra-class correlation coefficient of 0.80 or below, retaining 2,186 reproducible features. LASSO regression with cross-validation then selected the final feature set, capped at less than 10% of sample size to prevent overfitting. All four models (radiomics, morphological, quantitative, and combined) were trained using the Random Forest algorithm with grid-search hyperparameter optimization.
Morphological features for the conventional model included lesion size category, nodular pattern, lobulated sign, spiculated sign, bubble lucency, vascular sign, margin definition, bronchial distortion, pleural attachment, and pleural retraction. Quantitative features included 3D maximum diameter, mean CT value, volume, and mass. Interobserver agreement for morphological features exceeded a Cohen's kappa of 0.80 for all categories.
17.9% of residual nodules grew within 5 years. Of the 603 residual nodules, 108 (17.9%) were categorized as grown and 495 (82.1%) as stable. Among grown nodules, 95.4% showed size increase, 34.3% showed solid component growth, and 4.6% developed a new solid component.
Growth rates varied dramatically by size category: only 4% of nodules smaller than 5 mm grew, compared to 8.4% of 5-8 mm nodules and 48.5% of nodules larger than 8 mm. Part-solid nodules showed a 74.3% growth rate, while pure ground-glass nodules grew in only 14.4% of cases -- a striking difference that confirmed solid components as a major risk marker.
Morphological features significantly enriched in the growth group included spiculated sign, bronchial distortion, bubble lucency, pleural retraction, and pleural attachment. Quantitative parameters were also strikingly different: grown nodules had substantially larger 3D maximum diameters, higher mean CT values, and greater volume and mass than stable nodules.
Among the 79 patients with growing residual nodules, 20 (25.3%) had multiple growing nodules simultaneously, and 19 patients ultimately required additional surgical resection. Two patients developed lymph node metastasis during follow-up due to residual nodule progression, underscoring the clinical consequences of missed high-risk nodules.
Radiomics achieved the highest AUC in the test set. In the test set, the radiomics model achieved an AUC of 0.892 with an accuracy of 86.7%, outperforming the morphological model (AUC 0.834, accuracy 81.0%) and the quantitative model (AUC 0.862, accuracy 78.1%). The combined model achieved an AUC of 0.887, statistically equivalent to the radiomics model alone (p = 0.728).
Net reclassification improvement analysis confirmed the radiomics model's clinical superiority: it improved risk classification by 7.4% over the morphological model (p = 0.017) and by 14% over the quantitative model (p = 0.005). This quantifies not just statistical but actionable clinical improvement in identifying patients who need intensified management.
Calibration curves showed that the radiomics model had the best alignment between predicted probabilities and actual growth outcomes in the test set, with a Brier loss of 0.114. Decision curve analysis demonstrated positive net benefit for the radiomics model across probability thresholds of 0.1 to 0.6, while morphological and quantitative models showed minimal or negative benefit at lower thresholds.
The top-ranked radiomics feature was the major axis length of the nodule shape, reflecting the intuitive relationship between size and growth risk. Fourteen texture-based features captured spatial heterogeneity patterns associated with histopathological complexity and malignant transformation, providing information beyond what morphological inspection alone can reveal.
Small nodules lack visible morphological detail. Most residual nodules in this study were small or subcentimeter lesions with limited visible morphological information. This structural paucity likely explains why the morphological model underperformed compared to radiomics, which can still extract texture and shape information even from small nodules.
In the combined model, LASSO regularization eliminated all quantitative features due to multicollinearity with radiomics features, while retaining three morphological features -- vacuole sign, nodular pattern, and pleural adhesion. This suggests that certain clinician-identified morphological characteristics provide unique information not captured by radiomics, but conventional quantitative measurements are largely redundant.
The comparable performance of the radiomics and combined models reinforces radiomics as a standalone sufficient tool for clinical risk stratification. Future studies may explore deep learning architectures that could better integrate radiomics with complementary morphological data.
The decision curve analysis framework was adapted to the clinical context: rather than treating all patients the same, it compared the choice between concurrent surgical resection versus intensified surveillance, matching real-world decision points. The radiomics model's sustained net benefit across a wide threshold range directly informs these surgical planning discussions.
Radiomics as a standalone clinical tool. CT-based radiomics demonstrated superior and standalone performance for predicting 5-year growth of residual nodules after dominant tumor resection, offering high accuracy, robust discrimination, excellent calibration, and meaningful clinical net benefit without requiring any additional imaging or clinical data.
Key limitations include the retrospective single-center design, potential selection bias for small or anatomically obscured nodules, and the 5-year follow-up horizon that may not capture ultra-long-term behavior. Extended follow-up studies and serial radiomic assessments are needed to evaluate feature stability and model validity over longer periods.
The biological interpretability of the selected radiomics features requires validation through multi-omics correlation studies. Prospective research combining radiomic phenotyping with genomic sequencing of residual nodules could elucidate how imaging features map to molecular progression mechanisms.
Future directions include exploring deep learning architectures for improved multi-modal feature integration, multicenter validation, and potentially incorporating liquid biopsy approaches as complementary biomarkers to overcome the challenge of low tissue sampling rates from residual nodules.