JMIR Publications

ABSTRACT

Background:

Routine computed tomography (CT) can quantify whole-tumour features. Clinical use requires reproducibility across scanners and populations.

Objective:

To compare five CT-based algorithm categories for non-small-cell lung cancer genotyping using studies that evaluated multiple categories within shared patient cohorts.

Methods:

PubMed, Embase, Web of Science, and the Cochrane Library were searched to 25 June 2026. Eligible studies compared at least two categories against a shared molecular reference standard. A Bayesian diagnostic accuracy network meta-analysis synthesised epidermal growth factor receptor (EGFR) results. The Prediction model Risk Of Bias ASsessment Tool for artificial intelligence assessed bias. Certainty was evaluated using the Grading of Recommendations Assessment, Development and Evaluation framework. The review was funded by the Hunan Provincial Natural Science Foundation and two Hunan pharmacy research programmes.

Results:

Twenty retrospective studies were included. Seventeen EGFR studies contributed 43 validation datasets and 2450 validation patients, counted once per study. Convolutional neural networks had the highest sensitivity (0.84; 95% credible interval [CrI], 0.78–0.88). Transformers had the highest specificity (0.91; 95% CrI, 0.85–0.96), based on two datasets with 219 participants. Their diagnostic odds ratio relative to convolutional networks was 1.94 (95% CrI, 0.78–4.80). Certainty was moderate at most. Three studies of other genotypes were synthesised narratively.

Conclusions:

Convolutional neural networks and transformers showed complementary sensitivity and specificity profiles. Their comparative accuracy remained uncertain, and the evidence did not support replacing molecular testing. Clinical Trial: PROSPERO CRD420261479713; https://www.crd.york.ac.uk/PROSPERO/view/CRD420261479713