Lung cancer is the most common malignant tumor and the leading cause of cancer-related mortality worldwide (1). Non-small cell lung cancer (NSCLC) constitutes approximately 85% of all lung cancer cases, with adenocarcinoma being the most prevalent subtype (2). The widespread adoption of low-dose computed tomography (LDCT) for lung cancer screening has substantially increased the detection of pulmonary nodules, particularly subcentimeter subsolid nodules (SSNs) (3, 4). SSNs, encompassing both pure ground-glass nodules and part-solid nodules, exhibit heterogeneous biological behaviors, ranging from pre-invasive lesions (e.g., atypical adenomatous hyperplasia) to invasive adenocarcinomas (IAC) (5).

Accurately characterizing these nodules is critical, as stage I NSCLC (including stage IB disease) demonstrates remarkable heterogeneity in prognosis. Even after complete resection, the 5-year overall survival for stage IB NSCLC is approximately 73%, with recurrence rates as high as 18-29% (6, 7). Accurate preoperative prediction of invasiveness in SSNs remains a diagnostic challenge, as their imaging characteristics frequently overlap across different pathological subtypes. Conventional radiological assessment depends on subjective morphological characteristics such as size, margin, and solid component proportion, which often suffer from limited diagnostic precision (8). Furthermore, the role of adjuvant therapy in early-stage NSCLC like stage IB remains a subject of debate (9), highlighting the need for better tools to identify high-risk patients who might benefit from adjuvant therapy or more extensive resection.

Radiomics, which is an emerging field in quantitative imaging analysis, enables high-throughput extraction of minable data from medical images, capturing subtle patterns imperceptible to the human eye (10). When combined with machine learning (ML), radiomics facilitates the development of robust predictive models for nodule characterization (11). Recently, studies have endeavored to predict the invasiveness of subsolid nodules using advanced methodologies. For instance, Zuo et al. developed a combined nomogram integrating deep-learning-assisted CT texture features, achieving robust performance (12). Similarly, Li et al. demonstrated the superior performance of a stacking ensemble machine learning model that combines radiomic signatures with clinical-radiological features (13). These studies underscore the potential of quantitative imaging analysis. However, the specific challenge of preoperatively predicting invasiveness in SSNs, which are particularly elusive on conventional CT, remains less explored. Furthermore, the comprehensive integration of explainable AI (XAI) techniques, such as SHAP, is still needed to provide transparent, biologically plausible explanations for model predictions in this specific context (14). Therefore, this study aims to develop an interpretable radiomics-machine learning model specifically for SSNs, with a strong emphasis on leveraging SHAP analysis to elucidate the model’s decision-making process and enhance clinical trustworthiness.

Our study addresses two critical gaps in the radiomics literature on pulmonary nodules. First, we specifically targeted the diagnostic challenge of SSNs—a population increasingly detected by LDCT screening yet notably underrepresented in prior research, which often prioritized larger, solid lesions (16, 17). The management of these minute nodules remains highly controversial due to the limitations of subjective CT assessment (18, 19), creating a pressing unmet clinical need. Second, and more importantly, we moved beyond the common “black-box” paradigm by making model interpretability a cornerstone of our approach. Through the comprehensive integration of SHAP analysis, we not only predict invasiveness but also elucidate the underlying reasoning by identifying the most predictive features and explaining their directionality and potential biological significance. This commitment to transparency is a critical step toward building clinical trust and facilitating the integration of artificial intelligence tools into real-world decision-making processes (20, 21).

Our findings directly address the limitations of conventional CT assessment for SSNs (16, 22, 23). Although qualitative features such as vacuoles and vascular convergence showed statistical association with invasiveness in univariate analysis (Table 1), their diagnostic consistency was inadequate across cohorts. Radiomics objectively quantified an otherwise imperceptible tumor heterogeneity, yielding significantly improved diagnostic precision. The robust performance of the LR model (AUC 0.842) underscores radiomics’ potential to augment subjective evaluation, particularly for nodules displaying equivocal morphological features. The LR model in our study demonstrated balanced and robust performance (AUC: 0.842) in the internal validation, a result comparable to the advanced models reported for larger SSNs, such as the deep-learning-assisted nomogram by Zuo et al. (12) and the stacking ensemble classifier by Li et al. (13). This indicates that a well-constructed radiomics model can achieve high diagnostic accuracy without necessarily resorting to complex architectures. A critical finding from our statistical comparison was that DeLong’s test confirmed no significant difference in AUC between the top-performing LR model and the RF or SVM models. This statistical equivalence is pivotal, as it shifts the model selection criteria from pure discriminatory power to a holistic evaluation of generalizability, stability, and clinical applicability. Consequently, the selection of the LR model is strongly justified by its optimal balance of three factors. First, it exhibited superior generalizability, evidenced by the minimal AUC decline from training to validation (ΔAUC 0.031), which was relatively less than that of RF (0.089) or SVM (0.068), demonstrating robust resistance to overfitting (24, 25). Second, it achieved a clinically essential equilibrium between sensitivity (0.792) and specificity (0.733), effectively mitigating the risks of both under-treatment and unnecessary intervention, as further validated by its high F1 score (0.745) (26). Finally, and crucially, unlike complex “black-box” models, LR’s linear architecture provides an inherently transparent foundation (27), which we seamlessly integrated with SHAP analysis. This integration moves beyond mere prediction to explain the “why” behind each decision, providing both global and local explanations that are fundamental for building clinician trust and fostering clinical adoption (28). This combination of non-inferior predictive power, proven stability, and inherent explainability makes LR the most clinically pragmatic choice for deployment. A notable strength of our study is the implementation of hardware-level standardization for the external validation. By selecting an external cohort scanned with identical CT models and default parameters, we successfully isolated the biological and pathological variations from potential technical artifacts. The stable performance of the LR model across both centers suggests that our radiomics signatures are intrinsically linked to the tumor’s invasive potential rather than specific imaging noise.

The integration of SHAP analysis, a core novelty of our work, yielded unprecedented insights into the biological mechanisms captured by our model. The top-ranked features suggested potential links to tumor pathobiology: wavelet_HLL_glszm_LowGrayLevelZoneEmphasis: This feature quantified the spatial clustering of low-attenuation voxels in the wavelet-transformed HLL (horizontal-high, vertical-low, diagonal-low) sub-band. Higher values indicated larger contiguous regions of low CT attenuation (-600 to -800 HU). In IAC, aggressive tumor behaviors induced necrotic zones from rapid growth exceeding angiogenesis and microcystic fusion due to parenchymal collapse. These processes created macroscale low-attenuation clusters that were visible in the HLL sub-band, which emphasized horizontal structural disruptions. Consequently, elevated feature values correlated with advanced invasiveness (29).

Original_shape_Flatness: this radiomic feature measured the three-dimensional (3D) planarity or flatness of an ROI, with values on a scale of 0 to 1. Higher values (approaching 1) indicated a more planar or isotropic shape, geometrically approximating a cube or sphere. Conversely, lower values (approaching 0) corresponded to a more elongated, rod-like morphology. This feature primarily described the 3D geometric properties of the ROI, specifically its spatial directional dimensionality and deviation from sphericity. This may reflect the ROI’s growth pattern and interactions with surrounding tissues.

Evidence indicates that morphological features, including flatness, serve as effective differentiators between lung adenocarcinoma pathological subtypes (AIS/MIA vs. IAC) (30–32). The characteristically more irregular morphology of IAC, stemming from its aggressive growth patterns (destructive growth, complex structures such as micropapillary/solid components, and infiltration along structures), could manifest through alterations in flatness: (1) a decrease in overall flatness (suggesting a trend towards elongation), and (2) localized increases in flatness in specific areas (e.g., flat patches along the pleura). Critically, an overall increase in morphological complexity represented a more critical marker of invasiveness.

The log_firstorder_log.sigma.4.0.mm.3D_Minimum: this feature represented the minimum intensity value after Laplacian of Gaussian (LoG) filtering at σ=4mm scale. Lower values (more negative) indicated extremely low-attenuation foci (~-900 to -1000 HU). In pre/minimally invasive lesions (AIS/MIA), preserved alveolar airspaces generate isolated voxels with near-air density, and “bubble-like” pseudo-cavitations formed due to lepidic growth without destruction. The LoG filter (σ=4mm) amplified these submillimeter air-trapping zones. As invasiveness increased, tumor cells filled airspaces, reducing extreme low-attenuation voxels. Thus, lower minimum values (more negative) predicted lower invasiveness. Consistent with the findings of Zuo et al. (12). and Li et al. (13)., our SHAP analysis also identified features related to tumor morphology and patient demographics as top contributors. Specifically, the importance of features like wavelet_HLL_glszm_LowGrayLevelZoneEmphasis (potentially indicating necrosis) and original_shape_Flatness (reflecting morphological irregularity) in our model aligns with the principle that radiomics captures critical tumor heterogeneity. The concordance in key predictors, such as morphological complexity, across independent studies strengthens the biological plausibility of radiomic models. These SHAP-derived insights transcended simple feature lists, elucidating how specific quantitative image characteristics mechanistically contributed to invasiveness prediction. Future histopathological correlation studies linking these features to specific tumor microenvironment characteristics (e.g., fibroblast proliferation, architectural distortion, microvascular patterns) are essential.

The “black box” nature of numerous ML models remains a significant barrier to clinical adoption (15). Our integration of SHAP explicitly addressed this limitation. Global summaries provided clinicians with key radio-phenotypic determinants of SSN invasiveness, potentially refining human pattern recognition. Local explanations using force plots personalize predictions for individual patients and visually delineate the contribution of each feature towards the nodule-specific risk score. This transparency facilitated clinician-patient communication, supported tailored clinical decision-making (e.g., surgical planning versus active surveillance), and cultivated trust among radiologists and clinicians (33, 34). Furthermore, DCA validated the model’s net clinical benefit over a broad range of threshold probabilities, highlighting its practical advantage over binary treat-all or treat-nothing strategies. In terms of methodological rigor, our study design adheres to the core recommendations of the Radiomics Quality Score (RQS) framework (detailed in Supplementary Table 1). By implementing rigorous feature stability screening (ICC > 0.75), multi-step dimensionality reduction, and, importantly, independent external validation, we have endeavored to ensure the reproducibility and clinical relevance of our radiomic signature.

This study has several limitations that warrant consideration. First, although we successfully incorporated an independent external validation cohort, the retrospective nature of the study may still introduce selection bias. Since the model was developed and validated using data from a single institution with specific CT scanners and imaging protocols, the performance and stability of both the radiomic features and the predictive model may vary when applied to external cohorts with different acquisition parameters and demographic characteristics. Second, although SHAP analysis improves interpretability, the biological plausibility of the top radiomic features—while informed by existing literature—remains hypothetical and requires further validation through histopathological correlation. Future studies incorporating precise region-of-interest matching between CT imaging and histopathological samples (e.g., via tissue microarrays or detailed pathologic mapping) are necessary to confirm the underlying biological mechanisms of these image-based biomarkers. Third, to ensure model parsimony and mitigate overfitting risks given the sample size, clinical variables (e.g., age, smoking history) and semantic radiological features were not incorporated. Their inclusion in future iterations may improve model comprehensiveness and predictive power. Third, due to the retrospective nature of the data collection, we were unable to explicitly evaluate the performance differences between the specific CT scanner models (GE and Philips) used in our primary cohort. Although we applied rigorous IBSI-recommended image preprocessing (e.g., 1×1×1 mm³ spatial resampling and gray-level normalization) to minimize inter-scanner variability prior to feature extraction, future prospective studies should track specific hardware parameters to conduct comprehensive inter-vendor subgroup analyses. Finally, although logistic regression was selected for its favorable performance and interpretability, more complex yet explainable approaches—such as explainable boosting machines or deep learning-based radiomic models—may capture subtler patterns and should be investigated in larger, multi-center datasets.

While our study has been strengthened by the inclusion of an independent external validation cohort from a second center, it remains a retrospective analysis. Future prospective, large-scale multicenter studies are still warranted to further confirm the model’s robustness and clinical utility across more diverse clinical settings and imaging platforms prior to widespread implementation.

This study establishes a robust and interpretable radiomics-LR model for invasiveness prediction in subcentimeter SSNs. By combining rigorous feature extraction, ML validation, and SHAP explanations, this tool provides biological insights and clinical decision support. Prospective multi-center validation is warranted for clinical translation.