Deep Learning-Based Lung Cancer Classification from CT Images: A Systematic Review of Morphological Features, Model Architectures, and Classification Performance

Abstract

Context: Lung cancer accounts for approximately 12.4% of new cancer diagnoses and 18.0% of cancer-related deaths worldwide, and most patients present at an advanced stage. Computed tomography (CT) is the primary modality for lung cancer screening and characterisation; however, manual interpretation varies and is increasingly burdened by workload. Deep learning offers automated, high-accuracy classification. This systematic review synthesises evidence on deep learning-based lung cancer classification using CT images, focusing on morphological features, model architectures, and classification performance. Evidence Acquisition: This review followed the PRISMA 2020 guidelines. Four databases were searched on 5 July 2026, covering 2014 to 2026: Semantic Scholar via Elicit (n = 1,000), PubMed/MEDLINE (n = 1,204), IEEE Xplore (n = 421), and Scopus (n = 3,781). Eligible studies were peer-reviewed articles that applied deep learning with explicit morphological features (shape, margin, texture, density) to CT-based classification, used human CT datasets with a minimum of 50 CT examinations (the smallest included cohort comprised 96 patients), and reported at least one quantitative metric. Records were screened in Elicit and then independently double-checked by two human reviewers, the first author (M.T.) and the second author (S.H.Y.), at both the abstract- and full-text screening stages and during data extraction. Conference proceedings and preprints were excluded. Results: Twenty-seven studies met all eligibility criteria. All included models used convolutional neural networks; ResNet or 3D-ResNet backbones were the most common (10 studies), whereas transformer or state-space components appeared in four recent (2024 to 2026) studies. Morphological information was handcrafted or radiomics-derived in 14 studies and learned end-to-end in 13; the most frequently used feature groups were texture, shape and margin signs (spiculation, lobulation), density, and lesion size. Private datasets predominated (approximately two-thirds of studies); LIDC-IDRI or LUNA16 were used in roughly a quarter. Accuracy ranged from 66.3% to 99.9% (median 89.7%; 25 studies), and AUC ranged from 0.71 to 0.99 (median 0.92; 21 studies). Because this accuracy range combines binary and multiclass tasks with different chance levels and class balance, performance is reported by task rather than as a single figure, and sensitivity and specificity are prioritized over accuracy where studies allow. Only 10 studies (37%) reported external validation, and 8 (30%) treated metastatic (secondary) lesions as a distinct class.