Abstract
Early detection of lung cancer is hindered by tumour heterogeneity and cross-site variability in computed tomography (CT) acquisition, which often limits the generalisability of diagnostic models. This study aimed to develop and rigorously evaluate a generalisable deep learning–radiomics fusion framework for early detection across multi-centre datasets. Retrospective secondary CT data from the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset (≈1,010 subjects) and the Non-Small Cell Lung Cancer (NSCLC) Radiogenomics dataset (≈211 subjects) were curated and subjected to comprehensive quality control. Nodule-centred 2.5D (pseudo-3D) crops, combining adjacent CT slices to capture limited spatial context, were used to derive 512-dimensional image embeddings, which were concatenated with 93 handcrafted, non-shape radiomic descriptors extracted within expert-annotated masks, producing a 605-feature representation (512 + 93) per sample. Logistic regression, random forest, gradient boosting, and a compact fusion network were trained and evaluated under a strict leave-one-site-out (LOSO) validation design, with all preprocessing steps including median imputation, z-score normalisation, variance filtering, and threshold calibration fitted exclusively on training folds to prevent information leakage.
Across sites, discrimination was near-ceiling (area under the receiver operating characteristic curve (ROC-AUC) > 0.99; area under the precision–recall curve (PR-AUC) > 0.98). On the held-out LIDC site, the radiomics baseline achieved AUC = 0.9935 and F1 = 0.938 at a 0.5 threshold, while the fusion model reached AUC = 0.9824 and F1 = 0.928, at development-selected thresholds, both preserved precision ≈ 1.00, with a recall trade-off driven by the fusion model’s more conservative operating point. On the held-out NSCLC dataset, both approaches operated at ceiling performance (AUC ≈ 1.00; F1 ≥ 0.98; recall = 1.00). Macro-averaged across centres, the fusion attained ROC-AUC = 0.9902, F1 = 0.954, and accuracy = 0.953. Feature analyses revealed recurrent latent factors (e.g., f232, f279, f358, f511), alongside influential radiomics such as GLSZM_SmallAreaLowGrayLevelEmphasis, GLCM_MCC, and NGTDM_Busyness, indicating complementary representations. Overall, the fusion framework maintained consistent cross-site performance, mitigated domain shift, and provided an interpretable, scalable, and reproducible pathway for earlier detection, while motivating prospective, multi-institutional validation to confirm durability in broader clinical settings.