Artificial Intelligence-Assisted Lung Tumour Detection: A Systematic Review and Meta-Analysis

Yongming Zheng, Shichao Zhang, Lei Geng, Jiajing Li, Kaimin Shi, Di Yao, Qiong Li, Tao Zhang, Yusheng Zhao, Lei Zhang, Yi Sun, Jinxin Wan2025 · DOI

Introduction

Lung cancer remains the leading cause of cancer-related mortality globally. The prognosis for patients depends heavily on early detection; however, traditional clinical interpretation of medical imaging is often subject to human factors such as fatigue, subjectivity, and varying levels of expertise. Artificial Intelligence (AI), specifically deep learning architectures, offers a transformative potential to automate and refine this diagnostic process.

Despite the surge in AI research, the field currently lacks a high-level synthesis to determine if these algorithms consistently outperform traditional clinical methods across diverse environments. This systematic review and meta-analysis protocol, registered under INPLASY202590109, establishes a rigorous framework to benchmark AI performance and identify the factors driving diagnostic success.

Methodological Framework

The study is designed to aggregate and synthesize data from the global research community using a multi-step filtering and statistical process.

Search Strategy and Study Selection

The research team is conducting a comprehensive search across seven major databases, including PubMed, Embase, and the Cochrane Library. To ensure data integrity, the inclusion criteria require that studies provide sufficient information to reconstruct a 2x2 contingency table (True Positives, False Positives, True Negatives, and False Negatives). Furthermore, all included studies must utilize a high-confidence "reference standard" for ground truth, such as histopathology (biopsy) or multidisciplinary clinical consensus.

Quality Assessment (QUADAS-2)

To prevent the inclusion of biased data, the researchers are employing the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool. This framework evaluates the risk of bias across four critical domains: patient selection, the index test (the AI algorithm), the reference standard, and the flow and timing of the study.

Statistical Meta-Analysis

Quantitative synthesis will be performed using Review Manager 5.4. The primary metrics for evaluation include pooled sensitivity, specificity, and overall diagnostic accuracy. A key technical feature of this protocol is the use of the $I^2$ statistic to quantify heterogeneity (variance) between studies. If $I^2 \geq 50\%$, the team will transition from a fixed-effects model to a more conservative random-effects model to account for the inherent "noise" in diverse study designs.

Subgroup Analysis and "Algorithm Families"

A central objective of this research is to move beyond treating AI as a monolith. The protocol outlines specific subgroup analyses to investigate how different "algorithm families" (e.g., various Convolutional Neural Network architectures or Transformers) perform relative to one another. Additionally, the team will analyze the impact of training databases, such as the LUNA16 dataset versus localized hospital data, to determine how training diversity influences clinical robustness.

Evaluation of Performance and Bias

The study will utilize Risk Ratios (RRs) with 95% Confidence Intervals to compare AI performance against traditional control methods. An RR > 1 will serve as the mathematical threshold for declaring AI superiority. To ensure the findings are not skewed by "file drawer syndrome"—the tendency for researchers to only publish positive results—the team will apply Egger’s and Begg’s tests to detect and account for publication bias.

Conclusion and Clinical Significance

While the exclusion of preprints and studies with insufficient documentation may limit the inclusion of the most recent experimental developments, this protocol ensures that the final meta-analysis is built on a foundation of high-quality, peer-reviewed evidence. By identifying which algorithm families are most "clinic-ready" and pinpointing why performance varies across settings, this study will provide a vital roadmap for the integration of AI into routine oncological practice.

Summary updated

BibTeX
@misc{zheng2025artificial,
  author = {Yongming Zheng and Shichao Zhang and Lei Geng and Jiajing Li and Kaimin Shi and Di Yao and Qiong Li and Tao Zhang and Yusheng Zhao and Lei Zhang and Yi Sun and Jinxin Wan},
  title = {{Artificial Intelligence-Assisted Lung Tumour Detection: A Systematic Review and Meta-Analysis}},
  year = {2025},
  doi = {10.37766/inplasy2025.9.0109},
}
Figure