Prospective User Study and Multicenter Validation of Multimodal Medical Imaging Large Models

Background: Multimodal large models have shown significant potential in medical imaging. However, their performance and impact on clinical workflows across multiple centers require rigorous validation.

Objective: To assess the diagnostic performance of a multimodal large model and investigate whether AI assistance can improve the diagnostic accuracy and efficiency of radiologists with varying levels of experience.

Methodology: This research is designed as a multicenter, retrospective comparative reader study. A large-scale, diverse dataset of medical images (including CT and MRI) will be curated from the participating institutions. A group of licensed radiologists will perform diagnostic tasks in two separate sessions: a standalone session (without AI assistance) and an AI-assisted session, with a suitable washout period between sessions.

Data Analysis: The clinical "ground truth" will be established by expert consensus or histological results. The study will compare the Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, and specificity between the standalone and AI-assisted modes. Additionally, the reading time per case will be recorded to evaluate diagnostic efficiency.

Ethics: This study uses retrospective, anonymized data and does not alter the clinical management or treatment of patients.

The multimodal large model was developed and pre-trained using a massive dataset of approximately 1,000,000 medical imaging cases. This study focus on the multicenter clinical validation using an independent test cohort of 1,000 cases.

Inclusion criteria

• Patients who underwent systemic medical imaging examinations (e.g., CT or MRI) at participating centers for common systemic diseases.
• Imaging data must have confirmed clinical reference standards, expert consensus, or pathological diagnosis.
• Availability of complete DICOM format images with standard acquisition protocols.

Exclusion criteria

• Poor image quality (e.g., severe motion or metal artifacts) that precludes definitive diagnosis.
• Cases with incomplete clinical or pathological reference standards.
• Corrupted image files or duplicate cases.