A practical roadmap to implementation of in-house developed computational pathology algorithms

In the quiet, glass-walled rooms of a modern pathology lab, a new kind of helper has begun to work alongside the doctors who diagnose disease. For decades, the pathologist's job has been to examine thin slices of tissue under a microscope, looking for the tiny, telltale signs of cancer that the naked eye might miss. This work is slow, meticulous, and requires immense concentration. Recently, computers have learned to do this looking, too. Using a form of artificial intelligence, these programs can scan digital images of tissue slides and spot abnormalities with a speed and precision that rivals human experts. Yet, while the technology has advanced rapidly in research labs, it has rarely made the jump into the daily routine of a hospital. The gap between a computer program that works in a test and one that is trusted to help save lives in a real clinic is wide, filled with hurdles of safety, regulation, and the complex reality of how doctors actually work.

A team of researchers at the Radboud University Medical Center in the Netherlands has now bridged that gap. They did not just build a clever algorithm; they built a complete system to put that algorithm to work in their own hospital, right alongside their pathologists. Their goal was to solve a specific, high-stakes problem: determining if breast cancer has spread to the sentinel lymph nodes, the first nodes where cancer cells are likely to travel. In the past, doctors had to manually check every single slide, and if they didn't see cancer, they often had to order a second, more expensive test to be sure. The researchers wanted to see if their in-house computer tool could help them skip unnecessary steps, save time, and catch more cancer, all while following strict safety rules.

The journey began with a simple but difficult question: could a computer program be trusted to do this work safely? Before letting the tool touch a real patient's file, the team subjected it to a rigorous series of tests. They fed the program digital images of lymph nodes, including 48 negative sections and a total of 100 cases, some of which contained tiny clusters of cancer cells. They wanted to see if the computer would make mistakes, such as missing a cancer cell or falsely flagging a healthy spot as dangerous. The results were reassuring. When the computer looked at healthy tissue, it correctly identified it as negative one hundred percent of the time. When it looked at tissue with cancer, it found the vast majority of cases, missing only the very smallest clusters of cells. Crucially, the team also tested whether the computer would get confused by slight changes in how the slides were stained or scanned, mimicking the natural variations that happen in a busy lab. The tool remained steady and reliable, proving it could handle the messy reality of a working laboratory.

With the technical performance confirmed, the researchers moved to the next phase: seeing how the tool worked when real doctors used it. They set up a scenario where pathologists reviewed cases both with and without the computer's help. The results showed a clear benefit. When the doctors had the computer's assistance, they were able to spot cancer more often, catching cases they might have otherwise missed. More importantly, the tool changed the way the work flowed. In the old system, a doctor would look at a slide, and if they didn't see cancer, they would still have to wait for a second, specialized stain to confirm the result. With the new system, if the computer was confident that a slide was negative, the lab technician could immediately request that second stain without the doctor needing to look at the first slide again. This small change meant that for negative cases, the doctor did not have to spend time reviewing the initial image at all. For the cases where cancer was found, the time it took to reach a decision dropped significantly, from an average of nine seconds down to just four.

However, the team knew that speed and accuracy were not enough. In a hospital, safety is the absolute priority. Before the tool could go live, the team had to map out every single step of the process to find potential points of failure. They gathered pathologists, lab technicians, and safety experts to imagine everything that could go wrong, from a computer crash to a human error in reading the results. They found no major risks that could not be managed. They designed the system so that the computer's results were clearly displayed on the pathologist's screen, acting as a helpful guide rather than a replacement. If the computer flagged a spot as positive, the doctor would see it immediately. If it said negative, the technician would proceed with the next step, but the doctor remained in charge of the final decision. This careful planning ensured that the technology would support the doctors, not confuse them or bypass their expertise.

The final piece of the puzzle was fitting the tool into the existing digital world of the hospital. The researchers built a custom software bridge that connected their artificial intelligence to the hospital's main image system. This bridge acted like a secure courier, taking the digital slides from the doctors' view, sending them to the computer for analysis, and bringing the results back instantly. All of this happened within the hospital's own secure network, meaning no patient data ever left the building. The system was designed to be flexible, able to handle more work as the hospital grew, and simple enough that the staff did not need to learn a completely new way of working.

Once the system was live, the team watched how it performed in the real world. The tool is currently used in their pathology department, serving as a part of the workflow to help identify cancer more accurately and reduce the number of unnecessary tests that need to be run. The researchers found that the successful implementation of this tool required much more than just a smart algorithm; it required a deep understanding of how a hospital works, a commitment to safety, and a willingness to redesign the daily routine of the staff. They proved that when you build a system with the doctors and technicians in mind, and when you follow the strict rules of safety and regulation, artificial intelligence can move from a concept in a research paper to a trusted partner in the fight against disease. This work offers a clear path for other hospitals to follow, showing that the future of medical diagnosis is not about replacing the human expert, but about giving them the right tools to do their best work. While the tool is now in use, the team continues to monitor its long-term performance and safety through ongoing post-market surveillance, ensuring that its benefits are sustained over time.