A CT-based deep learning approach to differentiate multiple primary lung cancers, metastases, and benign nodules

Lung cancer, particularly adenocarcinoma and squamous cell carcinoma, remains a leading cause of cancer-related deaths globally. The diagnosis of multiple primary lung cancers (MPLCs) has become more frequent due to advanced chest CT technology and improved health surveillance. However, differentiating MPLCs from intrapulmonary metastases (IPMs) and multiple benign pulmonary lesions (MBPLs) remains challenging.

Distinguishing multiple primary lung cancers from metastases and benign lesions on CT remains challenging yet critical for treatment planning. Current approaches rely on subjective interpretation and invasive procedures. This study aims to develop and validate an automated deep learning classification system to provide rapid, objective diagnoses for optimizing patient management.

We studied 260 patients (MPLC = 83, IPM = 81, MBPL = 96; 881 axial CT slices). Six pretrained architectures (DenseNet-121, EfficientNet-B1, MambaOut-Kobe, ResNet-50, SwinV2-CR-Tiny-224, ViT-Tiny-Patch16-224) were compared in a five-seed ablation (seeds 42, 789, 1011, 2025, 2048). Pairwise one-vs-rest DeLong tests were aggregated across seeds to compare AUCs. Clinical utility was assessed using decision curve analysis (DCA). The final model (MambaOut-Kobe) underwent stratified five-fold cross-validation.

Considering efficiency, MambaOut-Kobe combined competitive accuracy with the lowest memory (~ 100 ± 14 MB) and low latency (~ 0.0093 ± 0.0017 s/image). Aggregated DeLong testing found no significant AUC differences among these models after multiplicity control. On five-fold cross-validation, MambaOut-Kobe achieved a macro-AUC of 0.946 ± 0.004 (95% CI 0.942-0.950), and an accuracy 0.829 ± 0.029 (95% CI 0.800-0.858). DCA demonstrated a positive net benefit across clinically relevant threshold probabilities compared with treat-all and treat-none strategies. Grad-CAM visualizations highlighted diagnostically relevant regions in CT images, providing interpretable decision-making support.

The MambaOut Kobe model demonstrates outstanding potential for clinical application in classifying MPLC, IPM, and MBPL. Its combination of high accuracy and computational efficiency makes it a promising tool for lung cancer diagnosis and treatment planning. This automated approach could reduce diagnostic uncertainty, minimize unnecessary invasive procedures, and facilitate timely, personalized treatment decisions for patients with multiple lung lesions. Future studies should focus on validating the model on larger, multicenter datasets and enhancing its discriminatory capacity between MPLC and IPM.