Study Purpose This systematic review examined how accurately different chest imaging techniques -- chest X-ray (CXR), CT scanning, and AI-assisted tools -- detect intrathoracic cancers (lung cancer, mesothelioma, lymphoma) in symptomatic patients presenting to primary care.
Search and Selection From an initial search identifying 30,539 records, 13 studies met all inclusion criteria. These studies evaluated patients with respiratory symptoms such as chronic cough, hemoptysis, or unexplained weight loss who underwent chest imaging in primary care or urgent care settings.
Key Findings CT consistently outperformed CXR for sensitivity. CXR sensitivity for lung cancer ranged from 33.3% to 75.9% across studies, while CT sensitivity ranged from 58% to 100%. An AI tool evaluated in one study underperformed multidisciplinary team (MDT) review with a very low positive predictive value of 5.6%.
Evidence Quality Quality was assessed using the QUADAS-2 tool for diagnostic accuracy studies and GRADE for certainty of evidence. Most included studies had moderate-to-high risk of bias due to selective patient populations and verification bias.
Wide Sensitivity Range Across the 13 included studies, CXR sensitivity for detecting lung cancer in symptomatic patients ranged from 33.3% to 75.9%. This wide range reflects differences in patient populations, radiologist experience, cancer prevalence, and cancer stage at presentation.
Specificity Considerations CXR specificity was generally higher than sensitivity, meaning the test is better at correctly identifying people who do not have cancer than at finding all who do. However, high specificity alone is insufficient when sensitivity is low in a high-stakes diagnostic context.
Normal CXR Does Not Rule Out Cancer A key finding is that a normal CXR does not reliably exclude lung cancer, particularly for small central tumors and early-stage disease. This has major implications for primary care practice where a normal CXR may prematurely end diagnostic workup.
Role of Radiologist Expertise Studies with specialist chest radiologist readers tended to show higher sensitivity than those using general radiologists. This highlights the importance of appropriate expertise when interpreting primary care chest X-rays for possible cancer.
Superior Sensitivity CT sensitivity ranged from 58% to 100% across studies, substantially higher than CXR. The upper range of near-perfect sensitivity was achieved in studies using dedicated low-dose CT screening protocols rather than standard diagnostic CT.
Detecting What CXR Misses CT detects small nodules, mediastinal involvement, and subtle parenchymal abnormalities that are invisible on CXR. For early-stage lung cancer -- where treatment is most effective -- CT provides a decisive advantage.
Barriers to CT in Primary Care Despite CT's superior performance, its widespread use in primary care is limited by cost, radiation exposure, equipment availability, and capacity constraints. Referral pathways to access CT are often slow in primary care systems.
Optimal CT Protocols Studies varied in whether they used low-dose or standard dose CT. Both performed well, but low-dose CT (LDCT) offers similar diagnostic accuracy with substantially lower radiation, making it preferable for repeated use or younger patients.
Single AI Study Evaluated Only one study among the 13 included evaluated an AI-assisted chest X-ray tool for cancer detection. This AI system was designed to flag CXR images with possible intrathoracic malignancy for expedited radiologist review.
Disappointing AI PPV The AI tool's positive predictive value (PPV) was only 5.6%, meaning 94.4% of AI-flagged cases did not have cancer. In contrast, MDT expert review achieved substantially higher PPV. This poor precision would generate enormous numbers of unnecessary referrals.
Context of AI Performance The AI's low PPV likely reflects the low prevalence of cancer in the broad symptomatic primary care population being screened. When applied to a population where most patients with cough do not have cancer, even a reasonably sensitive test generates many false positives.
Potential for AI Improvement The reviewed AI tool was an earlier-generation system. Newer AI tools incorporating clinical context alongside image analysis, and trained specifically on primary care populations, may achieve better precision without sacrificing sensitivity.
QUADAS-2 Assessment The QUADAS-2 tool evaluates diagnostic studies across four domains: patient selection, index test quality, reference standard adequacy, and flow/timing. Most included studies had at least moderate concern in one or more domains.
Common Bias Sources The most common methodological concern was selective patient recruitment -- studies often enrolled patients who were already suspected of having cancer, inflating apparent sensitivity compared to unselected symptomatic populations.
GRADE Certainty Applying the GRADE framework, evidence certainty for imaging performance estimates was rated low to moderate across outcomes. Sparse data, inconsistency across studies, and high risk of bias in most studies drove down certainty ratings.
Research Gaps The review identified a striking lack of high-quality studies evaluating imaging in unselected symptomatic primary care populations with cancer prevalence data representative of real primary care practice.
Primary Care Guidelines Current UK and European guidelines recommend urgent CXR for patients with lung cancer symptoms. This review suggests that a negative CXR should not terminate diagnostic workup for high-risk patients, and direct-access CT should be more widely available.
Optimal Imaging Pathways The review findings support implementing fast-track CT pathways for symptomatic patients with risk factors (smoking history, occupational exposure, hemoptysis) rather than requiring a normal CXR to precede CT access.
AI Integration Roadmap Before AI tools can be deployed in primary care cancer detection, they need training on representative populations, prospective clinical validation, and performance testing in the low-prevalence settings where they will be used.
Research Priorities Prospective studies of imaging in unselected symptomatic primary care populations, direct comparisons of CXR-first versus CT-first pathways, and formal evaluation of AI tools in primary care are the most urgent research needs identified.