MCP-MedImg: Enabling Standardized PACS-AI Integration through the Model Context Protocol

In the modern hospital, a patient's journey through medical imaging is a complex relay race. A technician captures a scan, such as a CT of the chest or an MRI of the heart, and the images are stored in a massive digital archive called a Picture Archiving and Communication System, or PACS. For decades, this system has been the central hub where doctors view and interpret these images. Recently, artificial intelligence has entered the race, offering powerful tools that can spot tumors, measure heart chambers, or flag abnormalities with incredible speed. However, connecting these AI tools to the hospital's existing archive has been a messy, custom-built process. Each new AI product often arrives as a standalone software package that requires its own installation, its own login, and its own way of talking to the archive. This creates a fragmented landscape where a radiologist might have to jump between three different programs to get three different AI opinions, and where the hospital struggles to keep a single, clear record of how those tools performed over time.

A new proposal called MCP-MedImg aims to untangle this knot by treating artificial intelligence not as a separate product to be installed, but as a service that can be called upon instantly and safely. The researchers behind this work, led by Xiong Shi, have designed a standardized set of rules that allows AI tools to plug directly into the hospital's existing archive without replacing the infrastructure. They built a system that automatically routes the right AI model to the right type of scan, hides sensitive patient information before the data leaves the hospital, and creates a complete, unbreakable record of every step the AI takes. This approach is designed to meet the growing demands of regulators who want to see how these tools perform continuously after they are approved, rather than just checking them once before they are sold.

The core of this work is a four-layer architecture that acts as a bridge between the hospital's data and the AI's brain. The first layer handles the physical movement of data, ensuring the system can talk to both modern web-based archives and older, traditional ones. The second layer is the security guard; it strips away names and identifying details from the images before they are analyzed, and it logs every single action with a cryptographic seal that proves the data has not been tampered with. The third layer is the traffic controller, which looks at the type of scan and the clinical question, then automatically dispatches the best available AI model to do the work. The final layer is the clinical supervisor, which decides whether a human doctor needs to check the AI's work. This decision is not a simple yes or no; it is a flexible rule that can require a doctor to review every high-risk case, allow them to skip the review for low-risk follow-ups, or let the system run automatically for urgent emergencies.

To prove this concept works, the researchers built a working prototype using open-source software and tested it with real medical data. They ran the system through two distinct challenges: analyzing 200 lung CT scans to look for nodules and processing 10 heart MRI scans to map the heart's chambers. In the lung tests, the system successfully identified nodules and generated standard medical reports in an average of 5.4 seconds, with even the slowest cases finishing in under 23 seconds. For the heart scans, which required a more complex analysis, the system took an average of 18.2 seconds to complete the task. Crucially, the system did not just produce results; it generated a detailed, 14-column log for every single scan, recording who asked for the test, which AI model was used, what the result was, and whether a human doctor reviewed it. This creates a transparent trail that regulators can inspect to ensure the tools are working safely and effectively over time.

The study also measured how much time the new system itself adds to the process. They found that the security checks, the routing logic, and the logging took less than one millisecond to complete. The vast majority of the time was spent simply moving the image data from the archive to the computer and running the AI calculation. This means the new system does not slow down the doctors; it simply adds a layer of safety and organization that was previously missing. The researchers demonstrated that the system could handle different types of scans, switch between different AI models automatically, and adapt its rules based on the urgency of the case. For example, in a simulated emergency where a doctor needed an immediate answer, the system could bypass the human review step to deliver results instantly, while still keeping a full record of what happened.

This work challenges the current way hospitals buy and use AI. Instead of treating each AI tool as a separate piece of software that must be installed and maintained individually, MCP-MedImg proposes a model where AI is a utility, like electricity or water, that flows through the hospital's existing pipes. The researchers showed that this is technically feasible and that it solves the problem of fragmented workflows. By standardizing how AI connects to medical images, the system allows hospitals to swap in new, better AI models as they become available without needing to rebuild their entire IT system. It also gives regulators the ability to monitor these tools in the real world, tracking their performance across thousands of cases to catch any errors or drifts in accuracy before they harm patients.

The researchers are clear that this is a prototype and that more work is needed before it can be used in every hospital. They tested it on specific types of scans, such as chest CTs and heart MRIs, and while the system is designed to handle other types of imaging, those have not yet been fully validated in this study. They also noted that while the system works well on a single powerful computer, running it across a large network of hospitals with thousands of simultaneous requests would require further testing. However, the results provide a solid foundation for the future. The system successfully demonstrated that it is possible to integrate artificial intelligence into medical imaging in a way that is secure, standardized, and fully auditable.

The ultimate goal of this research is to transform medical AI from a collection of isolated products into a governed public infrastructure. By creating a common language for how AI interacts with medical data, the researchers hope to lower the barriers for hospitals to adopt these tools, reduce the cost of integration, and increase the trust doctors have in the technology. The system ensures that every AI decision is traceable, every patient's privacy is protected, and every human override is recorded. This approach aligns with the new regulatory requirements emerging globally, which demand that medical devices be monitored continuously throughout their entire life cycle. The study suggests that by shifting from standalone software to a service-oriented model, the healthcare industry can finally realize the full potential of artificial intelligence, delivering faster, safer, and more reliable care to patients.