Spectral dual-layer detector CT-based radiomics-deep learning for predicting pathological aggressiveness of stage I lung adenocarcinoma: discrimination of precursor glandular lesions and invasive adenocarcinomas.

Background: Accurate diagnosis of early-stage lung adenocarcinoma (LA) subtypes is crucial for optimal patient management. Radiomics extract features from medical images reflect underlying biological information, while effective atomic number (Zeff) from new-generation spectral dual-layer detector computed tomography (SDCT) reflects tissue composition. This study evaluated the utility of SDCT-Zeff-based radiomics, deep learning (DL), and clinical features to differentiate between ground-glass nodule (GGN)-featured precursor glandular lesions (PGLs) and adenocarcinomas.

Methods: Patients diagnosed with GGN who underwent preoperative contrast-enhanced SDCT at two medical centers were prospectively enrolled between January 2022 and April 2024. Center 1 (Shengjing Hospital of China Medical University; n=582) served as the training cohort, while Center 2 (Shengjing Hospital, Huaxiang Branch; n=210) served as the external validation cohort. SDCT-Zeff delineated the region of interest (ROI) for radiomics feature extraction. A pre-trained ResNet50 model was used for DL feature extraction. Features were fused, screened, and integrated with various machine learning algorithms and clinical features to construct a clinical-based DL radiomics (DLR) signature nomogram, which was externally validated. Model performance was assessed regarding identification, calibration, and clinical utility.

Results: A total of 792 GGNs were analyzed, classified as glandular precursor lesions (n=296) and adenocarcinomas (n=496). Zeff was inversely correlated with invasiveness. Three features were obtained: clinical, radiomics, and DL. LightGBM was identified as the best-performing model. The area under the curves (AUCs) of DLR in the training and test sets were 0.974 [95% confidence interval (CI): 0.963-0.983] and 0.827 (95% CI: 0.770-0.884), outperforming radiomics (AUC =0.897 and 0.765), and DL (AUC =0.929 and 0.758). The nomogram coupling clinical features [Zeff_a, electron density (ED)_a, and tumor abnormal protein (TAP)] showed the best predictive ability, with AUCs of 0.983 (95% CI: 0.974-0.990) and 0.833 (95% CI: 0.779-0.885) in the training and test sets. The calibration curve indicated strong agreement between predicted and observed outcomes in both cohorts. Decision curve analysis (DCA) revealed that this nomogram offers significant clinical benefits, with a threshold probability range surpassing other models.

Conclusions: The coupled nomogram integrating SDCT-Zeff DLR with clinical features demonstrated improved predictive performance and was particularly effective in detecting GGN-featured glandular precursor lesions and adenocarcinomas. It provides a foundation for managing GGNs and offers valuable insights for preoperative evaluation.