Three distinct diagnoses, one imaging challenge. When a patient presents with multiple lung nodules or masses on CT, clinicians face a critical diagnostic question: are these separate primary lung cancers arising independently (multiple primary lung cancers, MPLC), spread from a single tumor site to different parts of the lung (intrapulmonary metastases, IPM), or entirely benign lesions (MBPL)? The answer determines whether treatment aims to cure each lesion separately, control systemic spread, or simply monitor a non-threatening finding.
Why accurate differentiation matters. MPLC and IPM have fundamentally different staging, prognosis, and treatment implications. Patients with MPLC may still be candidates for curative surgery on each lesion, while IPM indicates more advanced disease typically requiring systemic treatment. Misclassification can expose patients to surgery that offers no survival benefit or, conversely, deny surgical cure to patients who would benefit. Current diagnostic criteria require molecular profiling and pathological analysis that is time-consuming, invasive, and not always feasible.
CT features as non-invasive evidence. CT imaging captures morphological characteristics that differ between MPLC, IPM, and benign lesions - including nodule density (pure ground-glass, part-solid, solid), shape, margin characteristics, and spatial distribution within the lung. However, radiologists show significant inter-observer variability when interpreting these findings, particularly for distinguishing MPLC from IPM, which can appear visually similar.
Deep learning as a solution. Deep learning models can learn subtle image features from large training sets that are difficult to articulate as explicit rules. Applied to CT nodule characterization, these models have the potential to encode complex multi-feature patterns that distinguish MPLC, IPM, and MBPL more consistently than human readers, providing a non-invasive decision support tool.
Study population. The study enrolled 260 patients from Guangzhou, China with multiple pulmonary lesions confirmed by pathological diagnosis: 83 patients with MPLC, 81 with IPM, and 96 with MBPL. All patients had undergone CT imaging before treatment, and diagnosis was established using standard criteria including the 2013 IASLC guidelines for MPLC and pathological review of resected specimens.
Eligibility criteria. Patients were included if they had two or more pulmonary lesions on CT, pathological confirmation of lesion type, and complete clinical imaging records. Exclusions covered patients with prior lung surgery affecting imaging appearance, those on treatment that could alter lesion morphology before imaging, and cases where CT quality was insufficient for analysis.
CT acquisition and standardization. All CT scans were acquired on standard clinical scanners and underwent preprocessing to standardize image properties before deep learning analysis. Preprocessing steps included resampling to uniform voxel spacing, windowing to lung-appropriate Hounsfield unit ranges, and extraction of the dominant lesion region of interest for each patient to serve as the network input.
Five-fold cross-validation. To robustly estimate model performance despite the limited dataset size, a five-fold cross-validation framework was used. The 260 patients were divided into five subsets; each model was trained five times, each time using four subsets for training and the remaining one for testing. Final performance metrics represent averages and standard deviations across all five test folds.
Model selection rationale. Rather than relying on a single deep learning architecture, the study systematically compared six distinct model families representing different design philosophies: DenseNet-121, EfficientNet-B1, MambaOut-Kobe, ResNet-50, SwinV2-CR-Tiny-224, and ViT-Tiny-Patch16-224. This comparison allowed identification of the architecture best suited to this specific three-class CT classification task in terms of both accuracy and clinical deployability.
Architecture diversity. The six models span classic convolutional networks (DenseNet-121, ResNet-50), an efficiency-optimized convolutional model (EfficientNet-B1), a recent gated convolutional architecture (MambaOut-Kobe), a hierarchical vision transformer (SwinV2-CR-Tiny), and a pure patch-based vision transformer (ViT-Tiny). This diversity ensures the comparison captures whether convolutional inductive biases, attention mechanisms, or hybrid approaches are most beneficial for pulmonary lesion classification.
Training protocol. All models were initialized with ImageNet pre-trained weights and fine-tuned on the lung lesion dataset using identical training hyperparameters to ensure fair comparison. Standard data augmentation was applied during training to improve generalization. Each model was trained across all five cross-validation folds, and evaluation used the held-out fold for each iteration.
Evaluation metrics. Performance was assessed using area under the ROC curve (AUC), accuracy, and specificity. Decision curve analysis (DCA) was used to evaluate the clinical net benefit of each model across a range of decision thresholds - a metric that accounts for the relative costs of false positive and false negative errors rather than treating them equally as standard accuracy metrics do.
MambaOut-Kobe selected as optimal model. Across the five-fold cross-validation, MambaOut-Kobe achieved an AUC of 0.9415 and accuracy of 0.8195 in internal validation, with the best combination of performance and computational efficiency among all six architectures. The final model showed an accuracy of 0.829 plus or minus 0.029, AUC of 0.946 plus or minus 0.004, and specificity of 0.914 plus or minus 0.009 across the five folds.
Clinical efficiency advantages. A critical differentiator for clinical deployment is computational cost. MambaOut-Kobe required approximately 100 MB of memory - the lowest of any model tested - and processed each image in approximately 0.0093 seconds, enabling rapid real-time inference. These resource requirements make the model practical for hospital-grade computing infrastructure without specialized high-memory GPU hardware.
Decision curve analysis confirms clinical value. DCA showed that MambaOut-Kobe provided positive net clinical benefit across all three classification tasks (MPLC vs. others, IPM vs. others, MBPL vs. others) over a wide range of decision thresholds. This means that using the model's predictions to guide clinical decisions would result in better outcomes - fewer missed diagnoses and fewer unnecessary interventions - compared to treating all patients the same or using no model at all.
Comparative performance context. While other models showed competitive AUC values in some folds, none matched MambaOut-Kobe's combination of diagnostic accuracy, low memory footprint, and fast inference. The transformer-based models (SwinV2, ViT-Tiny) showed comparable or slightly lower AUC but substantially higher memory and inference time requirements, limiting their practical utility in routine clinical workflows.
Interpretability requirement. Deep learning models are often criticized as black boxes that produce predictions without explaining which image features drove the decision. In clinical settings, physicians need to understand why a model classified a lesion as MPLC rather than IPM to assess whether the prediction is trustworthy and consistent with anatomical knowledge. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to address this requirement.
What Grad-CAM shows. Grad-CAM generates a heatmap overlaid on the CT image showing which regions most strongly influenced the model's output for a given class. High-activation areas (shown in warmer colors) indicate the pixels the model weighted most heavily when making its classification decision. Ideally, these should correspond to the nodule itself and relevant surrounding tissue rather than irrelevant background regions.
Clinically relevant attention patterns. Grad-CAM analysis confirmed that MambaOut-Kobe consistently focused attention on the nodule itself and its immediate margins - the clinically relevant regions where morphological features distinguishing MPLC, IPM, and MBPL are most apparent. This alignment between model attention and clinician attention supports the model's validity and builds confidence that its predictions reflect genuine pathological differences rather than image artifacts.
Value for radiologist trust. The ability to show radiologists where the model is looking - and confirm it matches what they would examine - is important for clinical adoption. Heatmaps can also reveal cases where the model is attending to unexpected regions, flagging these for additional review and contributing to ongoing model improvement as edge cases are identified and corrected.
Addressing a high-stakes diagnostic gap. The study demonstrates that deep learning can meaningfully contribute to one of the more challenging diagnostic tasks in thoracic radiology. Distinguishing MPLC from IPM determines whether a patient is treated with curative or palliative intent, making even modest improvements in classification accuracy clinically significant. The model's strong AUC and specificity suggest it could serve as a triage or decision support tool to flag cases requiring multidisciplinary review.
Advantage over prior approaches. Previous deep learning work on pulmonary nodule classification has focused primarily on binary tasks (benign vs. malignant) or malignancy risk stratification. This study addresses a more clinically nuanced three-class problem that is directly relevant to staging and treatment planning. The inclusion of benign lesions as a third class reflects real-world clinical scenarios where not all incidentally found nodules are malignant.
Single-institution limitation. All 260 patients were enrolled from a single center in Guangzhou, China. CT scanners, acquisition protocols, and patient demographics may differ at other institutions and in other ethnic populations, potentially reducing model generalizability. External validation on data from multiple centers - ideally including Western populations - would be necessary before clinical deployment could be recommended.
MPLC versus IPM confusion. The most challenging classification boundary remains distinguishing MPLC from IPM, which sometimes requires molecular profiling of individual lesions rather than imaging alone. The model's performance on this specific subcomparison, while promising, reflects an inherent biological ambiguity that imaging features alone cannot always resolve. Future work integrating clinical variables, pathological biomarkers, or molecular data with CT features may further improve accuracy at this challenging boundary.