Vision-Language-Based Large-Scale Model for Emergency Screening

This single-center prospective randomized controlled trial adopted a parallel-group superiority design with a 1:1 allocation ratio. Randomization was performed via an independent central randomization system without stratification to avoid allocation bias. The trial is open-label for radiologists performing image interpretation, while outcome adjudicators and biostatisticians remain blinded to group assignment throughout analysis. Two study arms were set:

Experimental Arm: Radiologists interpreted thoraco-abdominal CT scans assisted by the powER vision-language foundation model. The model performs inference on input CT images and outputs ranked top 3 suspected critical lesions covering 18 categories of acute thoracic and abdominal emergencies, including aortic dissection, visceral lacerations, multiple fractures, acute pancreatitis, intestinal obstruction, etc.

Control Arm: The powER model also runs full background AI inference for every case and stores model outputs in the backend database, but no AI-generated lesion findings are displayed to the reading radiologists. Radiologists perform CT diagnosis on native PACS workstations relying solely on conventional clinical experience, with identical web interfaces as the experimental arm.

All eligible patients presented emergency-related clinical indications for thoraco-abdominal CT, with image quality meeting diagnostic standards. Pregnant women, patients with uncorrectable severe imaging artifacts, and those receiving only head or limb CT were excluded. All patient imaging and clinical data were fully de-identified for privacy protection.

The primary outcome measure was diagnostic decision-making time per case, defined as the time interval from the reporting radiologist opening the case on the web platform to the submission of diagnostic selections for target emergency lesions.

Secondary outcome measures included the ROC-AUC, sensitivity, specificity, PPV and NPV of lesion identification by reporting radiologists (using integrated clinical diagnosis, surgery and pathological results within 60-day follow-up as the reference gold standard), total report-writing turnaround time, report review turnaround time, radiologist diagnostic confidence score, and diagnostic consistency between reporting radiologists and the clinical reference standard (Kappa value).

Exploratory endpoints covered subgroup performance stratified by disease severity, identification capacity for rare emergency lesions, and auxiliary effects on radiologists with different seniority levels.

All participants signed written informed consent before enrollment. Standardized 60-day follow-up was implemented to record hospitalization duration, complications and all-cause 60-day mortality. Adverse events related to AI misdiagnosis or system failure were fully recorded and reported to the institutional ethics committee. All statistical analyses followed pre-specified statistical plans. The between-group comparison of the primary continuous time endpoint will use t-test or Mann-Whitney U test as appropriate. The DeLong test will be adopted for AUC comparison of secondary diagnostic outcomes, and mixed-effect regression models will be used for further adjusted analyses. The total planned enrollment will be determined based on the sample size calculation for the primary outcome of diagnostic decision-making time.