A Multirepresentation Model for Pulmonary Nodule Malignancy Risk Assessment and Model-Derived Risk Grouping on Routine Chest CT: a retrospective diagnostic modeling study

Technical Summary: A Multirepresentation Model for Pulmonary Nodule Malignancy Risk Assessment

Problem Statement
Accurate estimation of malignancy risk is critical for managing pulmonary nodules detected during routine chest CT. While low-dose CT screening has increased nodule detection, the majority are benign, creating a clinical challenge to reliably distinguish malignant from benign nodules at the initial evaluation. Existing management pathways rely heavily on size, attenuation, and follow-up intervals, while classical prediction models (e.g., Mayo, Brock) often fail to fully capture intranodular heterogeneity, global structural irregularity, and perinodular context from a single CT examination. Furthermore, reliance on repeated follow-up imaging increases costs, patient anxiety, and radiation exposure. Current deep learning and radiomics approaches often focus on specific subtypes, require extensive clinical context unavailable at initial assessment, or fail to explicitly model the complementary nature of internal texture, shape, and surrounding context.

Methodology
This retrospective, single-center study developed a TResNet-3D framework, a multirepresentation model designed to assess malignancy risk and derive risk groupings from routine chest CT.

• Data: The study utilized 18,914 pulmonary nodules from 3,172 patients (4,242 malignant, 14,672 benign). Data were split at the patient level (7:1:2) into training, validation, and independent test sets to prevent data leakage.
• Reference Standard: Malignancy was defined by pathology-confirmed primary lung cancer. Benignity was defined by benign pathology or at least 24 months of imaging stability without growth.
• Multiscale Feature Extraction:
  • Deep Features: Three cubic image patches (32³, 48³, and 64³ voxels) centered on each nodule were processed by a Taylor-enhanced 3D Residual Network (TResNet-3D). The architecture employed a shared shallow stem followed by scale-specific encoding branches (small for local texture, medium for global contour, large for perinodular context). A cross-scale attention mechanism reweighted streams, and outputs were fused into a 1,024-dimensional deep feature vector.
  • Radiomics: 74 candidate features were extracted using a standardized spherical Region of Interest (ROI) and an IBSI-oriented workflow (PyRadiomics). After filtering for variance and correlation, 35 features were retained.
  • Morphological Complexity: Seven low-dimensional geometric descriptors were computed from a pseudo-segmented foreground. Volume, surface-to-volume ratio, and sphericity were retained for modeling.
• Joint Modeling: The 1,024-dimensional deep features were fused with the selected radiomics and morphological features via a lightweight multilayer prediction head. This generated both a binary malignancy probability and a continuous risk score.
• Risk Grouping: Thresholds were fixed based on the validation set: Low risk (<0.3), Intermediate risk (0.3–0.7), and High risk (≥0.7).
• Evaluation: Performance was assessed using AUC, accuracy, sensitivity, specificity, calibration, decision-curve analysis (DCA), and gradient-based visualization (3D Grad-CAM). Comparators included ResNet18-3D, DenseNet69-3D, and traditional machine learning models (Logistic Regression, SVM, Random Forest).

Key Results
In the independent test set, the proposed TResNet-3D model outperformed all comparators:

• Discrimination: The model achieved an AUC of 0.9103 (95% CI 0.86–0.95), significantly higher than ResNet18-3D (0.8524) and DenseNet69-3D (0.8155).
• Classification Metrics: Accuracy was 0.8521, specificity 0.9318, and sensitivity 0.7059.
• Ablation Study: Deep features alone provided strong discrimination (AUC 0.8535). Radiomics alone performed poorly (AUC 0.6189). Simple fusion of deep and radiomics features did not improve performance over deep features alone. However, the addition of morphological complexity features to the deep-radiomics framework yielded the best performance (AUC 0.9103), indicating that low-dimensional shape descriptors provided critical complementary information.
• Risk Grouping: The model-derived scores showed a stepwise increase in observed malignancy rates across groups: Low risk (15.79%), Intermediate risk (42.86%), and High risk (82.55%).
• Clinical Utility: Decision-curve analysis indicated the model offered the highest net benefit across a broad range of threshold probabilities (0.35–0.85). Calibration curves showed reasonable agreement between predicted and observed outcomes, particularly in low-to-moderate risk ranges.
• Interpretability: 3D Grad-CAM visualizations confirmed the model focused on the nodule body, boundaries, and areas of irregular morphology or density heterogeneity, aligning with radiologic judgment.

Significance and Claims
The authors claim that the TResNet-3D framework provides a quantitative malignancy-risk estimation and model-derived score grouping from routine chest CT at initial evaluation. The study demonstrates that a multirepresentation strategy—integrating multiscale deep features with low-dimensional morphological complexity descriptors—can better capture the heterogeneity and structural irregularity of nodules than single-feature families or standard deep learning baselines.

The paper positions the model as an adjunctive decision-support tool rather than a standalone diagnostic replacement. The continuous score allows for descriptive risk grouping (low, intermediate, high), which may help stratify nodules into conservative assessment, clinical gray zones, or areas requiring closer diagnostic attention. The authors emphasize that while the model shows favorable discrimination and decision-analytic behavior, multicenter external validation and evaluation against standardized longitudinal endpoints remain necessary before clinical deployment. The work supports the transition from rule-based management to quantitative, image-derived risk modeling that preserves probabilistic outputs familiar to clinical decision-making.