en.Wedoany.com Reported - On May 6, 2026, the Beijing Academy of Artificial Intelligence (BAAI), in collaboration with Beijing Anzhen Hospital Affiliated to Capital Medical University and the First Affiliated Hospital of Henan Medical University, officially released the industry's first Cardiac Magnetic Resonance (CMR) multimodal diagnostic agent—BAAI Cardiac Agent. This system achieves a full-process diagnostic and treatment loop encompassing "structure segmentation and analysis—quantitative functional assessment—disease diagnosis and classification—intelligent reporting." It can perform end-to-end automated analysis of cardiac MRI images and output standardized reports that comply with clinical norms.
Cardiac magnetic resonance imaging is recognized as the "gold standard" in the diagnosis of cardiovascular diseases. With its characteristics of being radiation-free, multi-parametric, and high-resolution, it can clearly present details of cardiac structure, function, and myocardial tissue pathology. However, the clinical value of this technology has long been constrained by two major bottlenecks: First, image interpretation involves multi-sequence, multi-phase image analysis and precise calculation of multiple quantitative indicators, heavily relying on experienced specialist physicians, leading to long talent cultivation cycles. Second, there is a severe geographical imbalance in the distribution of high-quality diagnostic resources; top-tier CMR interpretation capabilities are highly concentrated in leading medical institutions in first-tier cities, while a large number of primary hospitals "can take images but cannot read them," making it impossible to replicate high-quality diagnostic capabilities at scale.
The core breakthrough of the BAAI Cardiac Agent lies in using AI technology to achieve the standardization, replicability, and stable, round-the-clock output of top-tier CMR diagnostic capabilities. The system is not a single AI model but an agent system based on an Agent-Expert architecture, capable of dynamically coordinating multiple "expert" sub-models. Its design is inspired by the collaborative model of human expert teams: a central multimodal agent orchestrates the overall process, while multiple specialized sub-models dedicated to cardiac structure segmentation, cardiac function quantification, myocardial tissue characterization, and disease diagnosis work in a coordinated division of labor. This end-to-end architecture allows it, much like a human expert team, to collaboratively complete the entire process from raw image upload, sequence identification, intelligent frame extraction, to structure segmentation, indicator calculation, disease diagnosis, and finally report generation, achieving a closed loop of "input images, output diagnosis." The diagnostic process efficiency is improved by approximately 30 times compared to manual work.
The research team comprehensively evaluated the BAAI Cardiac Agent on a CMR dataset of 2,413 patients from two partner hospitals, covering 7 major cardiovascular diseases, and the results were outstanding. In the internal validation set, the average AUC (Area Under the Curve) for disease diagnosis was 0.96; in the more challenging external validation set, the average diagnostic AUC reached 0.87, demonstrating strong cross-institutional generalization capability. The system can screen for heart diseases in a three-category classification of "normal/ischemic cardiomyopathy/non-ischemic cardiomyopathy" and can accurately differentiate non-ischemic cardiomyopathy into 5 subtypes, including hypertrophic, dilated, and inflammatory types. For core cardiac function indicators such as left ventricular ejection fraction, stroke volume, and left ventricular mass, the system's output results were highly consistent with manual measurement reports by clinical experts, with Pearson correlation coefficients all exceeding 0.90. The system compresses the manual interpretation process, which originally required 30 to 60 minutes, to about 1 minute, and reduces the impact of human experience variance on diagnostic results through a standardized analysis system.
In terms of report quality, the structured clinical reports automatically generated by the BAAI Cardiac Agent cover core content such as imaging manifestations, quantitative indicators, and diagnostic opinions, with standardized formatting and clear logic. After a blind review by 6 radiologists of varying seniority levels, the system's reports showed a high degree of consistency with reports written by experts. The system also supports downloading reports in PDF format, facilitating clinical archiving and subsequent diagnostic and treatment reference.
Alongside the product release, BAAI announced that the core code and supporting evaluation data for the BAAI Cardiac Agent have been fully open-sourced. They also simultaneously released the industry's first deep evaluation dataset for CMR image semantic understanding—CMRAgentEvalSet. This evaluation set covers the assessment of normal/abnormal mitral and tricuspid valves, comprehensive evaluation of left and right ventricular size, wall thickness, wall motion, and systolic/diastolic function, as well as detailed descriptions of pericardial effusion, perfusion status, and the three core sequences of CINE, LGE, and Rest MPI, providing a reproducible evaluation benchmark for academia and industry.
The Beijing Academy of Artificial Intelligence is a new-type research institution jointly established with superior units in Beijing's AI field, supported by the Ministry of Science and Technology and the Beijing Municipal Government. The release and open-sourcing of the BAAI Cardiac Agent targets the long-standing structural challenge of uneven distribution of medical resources in China—enabling patients in county-level hospitals and remote areas to access cardiac imaging diagnostic services of the same quality as those in tertiary hospitals, promoting the下沉 (downward allocation) and universalization of high-quality medical resources.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
