This study employed a prospective, randomised crossover trial design to evaluate the clinical utility of the TRACE artificial intelligence system for gastric cancer T-staging. A total of 54 radiologists from tertiary and non-tertiary hospitals, including both senior and junior practitioners, were enrolled. The study aimed to investigate whether AI-assisted diagnosis could improve the diagnostic accuracy of gastric cancer T-staging compared with independent interpretation by radiologists.
All participants were required to interpret 60 contrast-enhanced CT cases sequentially, completing two readings for each case: one without AI assistance and one with AI assistance; The order of the two readings was randomised, and a one-month washout period was observed between readings to eliminate memory bias. All cases were pathologically confirmed gastric cancer cases (stages T1-T4b), and the study simultaneously recorded the physicians' T-staging diagnostic results and the time taken per case. The 60 cases per radiologist were randomly selected from a pool of 1,000 histologically confirmed gastric cancer cases, stratified by pathological T stage T1-T4b. The reference standard was postoperative pathological T stage. The primary outcome was the change in T-staging accuracy between AI-assisted reading and standard (unaided) reading.The term "prospective" in this study refers to the prospective execution of radiologist enrollment, randomization, reading procedures, and data collection.
Contrast-enhanced CT (CE-CT) images of gastric cancer patients from the Liaoning Cancer Hospital;
Patients with a definitive postoperative pathological diagnosis of gastric cancer and a clear T-stage classification (T1-T4, including T4a and T4b);
Imaging data must be complete and of sufficient quality to meet diagnostic and analytical requirements, with no significant artefacts or missing key data;
Complete clinical and pathological information must be available to establish a diagnostic gold standard for comparison.
Severe missing imaging data or quality failing to meet analysis requirements (e.g., severe motion artefacts);
Lack of clear postoperative pathological T-staging results;
Cases not involving gastric cancer or with incomplete pathological information;
Cases of duplicate enrolment or inconsistent data recording.
Crossover
Utilizing the TRACE model to assist radiologists in T-staging
Diagnostic testAI-assisted reading: Radiologists interpret preoperative contrast-enhanced CT images for gastric cancer T staging with the support of the TRACE artificial intelligence decision system. The AI system provides a suggested T stage and relevant imaging features. The radiologist makes the final staging decision after reviewing the AI output. This intervention is used only during the AI-assisted reading session.
washout period
Other interventionParticipants are required to observe a washout period of at least 30 days between consecutive interventions/assessments.
Utilizing the TRACE model to assist radiologists in T-staging
Diagnostic testAI-assisted reading: Radiologists interpret preoperative contrast-enhanced CT images for gastric cancer T staging with the support of the TRACE artificial intelligence decision system. The AI system provides a suggested T stage and relevant imaging features. The radiologist makes the final staging decision after reviewing the AI output. This intervention is used only during the AI-assisted reading session.
Accuracy
Accuracy of radiologists' interpretation of T staging
Accuracy Change by Physician Experience Level
Changes in diagnostic accuracy of radiologists with different experience levels before and after AI assistance.
Stratified diagnostic accuracy of different T-stages
Stratified diagnostic accuracy for different T-stages (T1-T4, including T4a and T4b).
Agreement between physician diagnosis and pathological gold standard
Agreement between radiologists' diagnostic results and the pathological gold standard (e.g., Kappa value).
Agreement between AI model and physician interpretation
Agreement analysis between AI model prediction results and radiologists' interpretations.
Influence of case characteristics on AI assistance effect
Influence of different case characteristics (e.g., tumor location, size) on the performance of AI assistance.
Impact of individual physician differences on AI assistance effect
Impact of individual differences among physicians on the performance of AI assistance.
Value of AI assistance in reducing diagnostic discrepancy
Potential value of AI assistance in reducing diagnostic differences and improving reading agreement.
Impact of model probability information on physician decisions
Preliminary analysis of the impact of probability output from AI model on physician decision-making behavior.
Click on the lead sponsor to view all of their trials.
