Lung cancer remains a significant global health concern, with high mortality rates largely attributed to late-stage diagnosis. The diverse imaging and histological presentations of lung cancer pose challenges for clinicians in selecting the most effective treatment strategy. Early detection is critical for improving patient outcomes, but current methods often fail to identify small or subtle lesions. This necessitates the exploration of novel diagnostic approaches. Artificial intelligence (AI), particularly machine learning (ML), offers a promising avenue for improving the accuracy and efficiency of lung cancer diagnosis. ML algorithms can analyze medical images (e.g., X-rays, CT scans) and identify patterns indicative of malignancy, potentially aiding in earlier detection and facilitating more informed treatment decisions. This systematic review aims to evaluate the diagnostic accuracy of existing ML AI architectures in the detection and classification of lung cancer, providing a comprehensive overview of their current capabilities and limitations.

The introduction adequately sets the stage by highlighting the significant problem of late-stage lung cancer diagnosis and the need for improved diagnostic tools. It mentions the challenges posed by the disease's heterogeneity and the potential of AI to address these challenges. The literature review is implicitly integrated into the introduction, referencing several key studies on the use of AI in lung cancer diagnosis. While not explicitly separated into a distinct section, the introduction effectively summarizes relevant prior research on the efficacy of AI in this context.

This systematic review, conducted in February 2023, adhered to PRISMA and PROSPERO guidelines and was registered on the Open Science Framework (OSF). Four databases (PubMed, Web of Science, Cochrane, Scopus) were searched for English-language articles published up to December 2022, using relevant MeSH keywords related to lung cancer, AI, machine learning, and diagnostic imaging. The initial search yielded 5894 articles. After removing duplicates and screening abstracts, 315 full-text articles were assessed for eligibility, ultimately resulting in the inclusion of nine studies. Data extraction included study characteristics (author, country, year, design, quality assessment), findings (number of patients, AI architecture, comparison group, lesion types), performance metrics (TP, TN, FP, FN), and AI architecture specifics (sensitivity, specificity, accuracy). Included studies involved adult patients screened for lung cancer, with ML algorithms (neural networks and CADs) analyzing medical images for lung cancer detection. Studies using phantom images, non-imaging modalities, or assessing image segmentation without ML architectures were excluded. Two investigators independently evaluated the quality of included studies using the NHLBI's Study Quality Assessment Tools, with additional use of the Quality Assessment Tool for Observational Cohort and Cross-Sectional Investigations where appropriate. Studies were categorized as excellent (10+), good (5-9), or fair (0-4) based on scoring.

The analysis encompassed nine studies conducted between 2014 and 2022 across various countries, utilizing diverse ML architectures (ANN, EDM, PNN, SVM, POMDP, and RFNN) and imaging modalities (CT, HRCT, LDCT, X-rays, RADS). Study designs included case-control, retrospective cohort, and prospective cohort studies, with quality assessments ranging from excellent to fair. The number of patients varied significantly across studies (32–5402). Comparison groups varied, using methods like microscopic analysis, expert radiologist opinions, random X-rays, and random slices from healthy lung scans. Overall, the ML architectures showed promising results in detecting and classifying lung cancer. ANN, EDM, and SVM demonstrated effectiveness in detecting SCLC and NSCLC, while PNN, SVM, POMDP, and RFNN effectively differentiated malignant from benign lesions. Performance metrics (sensitivity, specificity, and accuracy) varied widely among studies. While some studies reported extremely high accuracy (100%), others showed lower accuracy (77.8%). These variations highlight the influence of factors such as study design, data quality, and the choice of ML architecture on the diagnostic performance of AI systems. Specific metrics such as sensitivity ranged from 0.81 to 0.99, while specificity ranged from 0.46 to 1.00 across the included studies. The variation in results emphasizes the need for more standardized methodologies and larger, more diverse datasets for future research.

The findings of this systematic review demonstrate the potential of ML AI architectures as valuable tools in lung cancer detection and classification. The diverse range of architectures and performance metrics across the included studies highlights the need for further research to standardize methodologies, optimize algorithms, and evaluate performance across diverse populations. The variations observed can be attributed to differences in study design, patient characteristics, imaging modalities, and ML architecture selection. Although some studies reported near-perfect accuracy, others showed considerable variability, underscoring the need for larger, well-designed, multi-center studies to establish the true clinical utility of these AI tools. Comparison with previous studies (Nasrullah et al., Ardila et al., Nam et al.) reinforces the promising potential of AI, with deep learning models showing high accuracy in some instances; however, traditional ML may be superior in scenarios with limited data. This discussion reinforces the ongoing need for development and validation of AI algorithms in diverse populations and the exploration of optimal combinations of imaging modalities and AI algorithms for enhanced lung cancer diagnosis.

This systematic review evaluated the diagnostic accuracy of ML AI architectures in lung cancer detection. The results indicate promising potential, but significant variability exists across studies. AI architectures demonstrated effectiveness in differentiating malignant from benign lesions and identifying specific lung cancer types. While promising, further research is needed to optimize ML architectures, increase the size and diversity of datasets, and validate findings in large-scale, multi-center trials. This optimization is essential to enhance the reliability and generalizability of these AI tools for real-world application in improving lung cancer diagnosis and patient care.

This review has several limitations. The heterogeneity of included studies (patient populations, imaging techniques, lesion types, ML architectures) may affect the generalizability of findings. The relatively small number of studies included and potential publication bias warrant caution in interpreting the results. The varied quality of the included studies, with some having small sample sizes or lacking methodological detail, could also impact the reliability of the findings. The lack of a pooled data analysis, due to the variability in methodologies and metrics, restricts the ability to draw definitive conclusions regarding overall diagnostic accuracy. Finally, the review focused solely on diagnostic accuracy, neglecting the broader clinical implications of AI in lung cancer management, including cost-effectiveness and impact on patient outcomes.