Abstract
<title>Abstract</title> <p>Background Accurate malignancy-risk estimation is central to pulmonary nodule management. Existing approaches do not always jointly capture intranodular heterogeneity, global structural irregularity, and perinodular context from a single routine CT examination. We developed a multirepresentation TResNet-3D framework for malignancy-risk assessment and model-derived risk grouping at initial evaluation. Methods This retrospective single-center study included 18,914 pulmonary nodules from 3,172 patients. Multiscale three-dimensional patches centered on each nodule were used to characterize internal texture, overall morphology, and perinodular context. A Taylor-enhanced 3D residual network extracted deep features, which were fused with radiomics and low-dimensional morphological complexity features to generate malignancy probability and a continuous score. Malignancy was defined by pathology-confirmed primary lung cancer, and benignity by benign pathology or at least 24 months of imaging stability. Data were split at the patient level into training, validation, and independent test sets. Performance was evaluated primarily by area under the receiver operating characteristic curve (AUC), with accuracy, sensitivity, specificity, calibration, decision-curve analysis, ablation experiments, and gradient-based visualization as secondary assessments. Results In the independent test set, the model achieved an AUC of 0.9103, an accuracy of 0.8521, a specificity of 0.9318, and a sensitivity of 0.7059, outperforming the radiomics model and the single 3D CNN comparators. Model-derived low-, intermediate-, and high-risk groups showed stepwise increases in observed malignancy rates. Decision-curve, calibration, and gradient-based visualization analyses supported the clinical plausibility of the model output. Conclusion The TResNet-3D provided quantitative malignancy-risk estimation and model-derived score grouping from routine chest CT at initial evaluation. These findings support further investigation as an adjunctive decision-support tool, although multicenter external validation remains necessary.</p>