Abstract Early detection of lung cancer remains one of the most effective strategies for improving survival; however, diagnostic performance is limited by variability in imaging protocols, scanner settings, lesion characteristics, and inter-reader interpretation. Deep learning (DL) models enable automated feature representation from medical images, while radiomics provides interpretable handcrafted biomarkers that capture tumour morphology, intensity, texture, and heterogeneity. When integrated into a clinical decision support system (CDSS), these approaches have the potential to augment radiologist decision-making, reduce false positives, and improve triage of indeterminate pulmonary nodules. However, both approaches continue to face challenges related to reproducibility, explainability, and generalisability across multi-centre imaging datasets. This systematic review followed PRISMA 2020 guidelines and synthesised evidence on CT-, LDCT-, spectral CT-, and PET/CT-based radiomics, DL, and DL–radiomics fusion models for early lung cancer detection and classification. Following revised searches across Scopus, PubMed/MEDLINE, IEEE Xplore, and Google Scholar, and subsequent eligibility screening, the retained core detection and classification studies were organised into three analytical categories: radiomics-only models, DL-only models, and DL–radiomics fusion models. Reported performance metrics, validation strategies, imaging modalities, fusion approaches, and centre designs were compared to identify methodological patterns and sources of heterogeneity. Fusion models generally outperformed radiomics-only and DL-only approaches across benign–malignant nodule classification, lung cancer classification, invasive adenocarcinoma prediction, ground-glass nodule assessment, and histological subtyping. Late and hybrid fusion strategies were more robust than early feature concatenation, particularly when handcrafted radiomics, deep features, and clinical variables (e.g., smoking history, spiculation, tumour stage, PET/CT indicators) were combined. However, the evidence base is constrained by retrospective designs, small samples, inconsistent segmentation and harmonisation, limited prospective validation, and incomplete reporting of calibration, decision-curve analysis, and external performance. Formal statistical pooling was not appropriate given the heterogeneity in outcomes, imaging protocols, feature pipelines, model architectures, and validation strategies. To address these gaps, this review proposes a methodological framework integrating multi-centre harmonisation, robust feature engineering, mathematically defined fusion strategies, domain generalisation, explainability, calibration, and clinical utility assessment to support reproducible and clinically meaningful CDSS for early lung cancer detection.
Tshivhula et al. (Thu,) studied this question.