Machine learning for differentiating primary lung cancer from pulmonary metastases and non-small cell lung cancer histological subtype classification: a systematic review.
Authors
Affiliations (3)
Affiliations (3)
- Department of Medical Physics and Biomedical Engineering, School of Medicine, Shiraz University of Medical Sciences, Shiraz, Iran. Electronic address: [email protected].
- Department of Medical Physics and Biomedical Engineering, School of Medicine, Shiraz University of Medical Sciences, Shiraz, Iran; Ionizing and Non-ionizing Radiation Protection Research Center (INIRPRC), Shiraz University of Medical Sciences, Shiraz, Iran.
- Department of Medical Physics and Biomedical Engineering, School of Medicine, Shiraz University of Medical Sciences, Shiraz, Iran; Nanomedicine and Nanobiology Reseach Center, Shiraz University of Medical Sciences, Shiraz, Iran.
Abstract
This study aimed to systematically review machine learning (ML) methods applied to computed tomography (CT) and positron emission tomography (PET)/CT for differentiating primary lung cancer from pulmonary metastases and for lung cancer histological subtype classification. PubMed/MEDLINE, Web of Science, and IEEE Xplore were searched for original peer-reviewed studies published from January 2010 to May 31, 2026. Eligible studies applied ML, deep learning, or radiomics to CT, PET, or PET/CT for primary versus metastatic lung lesion differentiation or major lung cancer histological subtype classification. Methodological quality was assessed with Quality assessment of comparative diagnostic accuracy studies (QUADAS-2), and reporting was evaluated using 10 prespecified Checklist for Artificial Intelligence in Medical Imaging-aligned items. Twenty-two studies were included after eligibility reassessment. Twenty-one provided a numeric study-level Area Under the Curve (AUC); the median representative AUC was 0.889 (IQR 0.820-0.936). The median AUC was 0.923 for primary/metastatic differentiation and 0.874 for histological subtype classification. Only 5 of 22 studies (22.7%) used independent external testing; among 4 externally validated studies with numeric AUCs, the median was 0.720. QUADAS-2 showed frequent uncertainty in patient selection (17/22, 77.3%), while CLAIM-aligned assessment identified limited external testing, code/model availability (3/22, 13.6%), and explainability reporting (4/22, 18.2%). CT- and PET/CT-based ML models show promising diagnostic performance for lung lesion-origin and histological subtype classification, but performance is lower in independent external evaluation. Broader multicentre validation, transparent reporting, reproducible code/model access, and prospective assessment are required before routine clinical implementation.