While every human being is unique, our basic anatomy has a common, predictable pattern of bones, muscles, ligaments, tendons and skin that identifies us all as homo sapiens. But nature isn’t always predictable: people can be born with a variety of uniquely shaped or even missing components that, while not incompatible with survival, can still make everyday life a challenge.
Concave chest cavities, atrophied calf muscles, jaw asymmetries, collapsed windpipes—ailments like these have plagued mankind throughout human history. But we’re now at a time when research and technology are coming together to provide astoundingly precise solutions to such ills. CT body scanning, data visualization and analysis software, highly accurate human-tissue modeling, and rapid manufacturing are converging to give patients customized replacements for what nature didn’t provide. Most recently, deep learning (a form of AI) is further accelerating the delivery of these medical solutions to meet growing patient demand.
A fascinating example of this convergence is the story behind the creation of AnatomikModeling, a service provider that helps doctors develop corrective prosthetics. The initial spark was a chance meeting between company founder Benjamin Moreno—at the time a neuroscience student with a background in bioinformatics—and a professor of Egyptology.
Applying bee brain work to mummies
Moreno was studying the brains of honeybees and mice, using confocal microscopy to identify individual neurons and their connections. “The idea was to visualize and quantify these neural networks, and find solutions in less time,” he says. While at university in Toulouse, France, he met a professor who wanted to use scanning technology to explore the interior of an Egyptian mummy housed in a local museum.
“I thought if we could scan a honeybee brain, we could certainly scale up and scan a mummy in order to explore its structure in a more scientific way,” says Moreno. The professor got permission to get the mummy scanned in a hospital CT machine (after regular patient hours). Moreno had heard about advanced CT-scan-data analysis and reached out to software provider Volume Graphics (now part of Hexagon). “They were very interested in my project and they like to support research and innovation,” says Moreno. He was given a free, one-year license of VGSTUDIO MAX software to study the mummy-scan data alongside the Egyptology professor, and subsequently founded his first company (IMA Solutions) to provide services to other institutions in the “heritage and cultural universe.”
A chance meeting creates a link to medicine
As his proficiency with the software progressed, Moreno (whose father’s experience as an orthopedic surgeon had piqued his interest in medicine) crossed paths with professor Jean-Pierre Chavoin, the head of plastic and reconstructive surgery at Toulouse University Hospital. “Prof. Chavoin’s idea was to create 3D custom-made implants for patients suffering from congenital deformities,” Moreno said. “This was very exciting to me. I knew we now had the analysis tools to help make this medical-related work a reality.” He and the surgeon began collaborating in 2008—resulting in the formation of the medical-focused design and engineering firm, AnatomikModeling.
Medical implants of many types have certainly been in use for years; think of such high-volume, routine procedures as breast enhancements (or repairs after cancer surgery), or joint replacements for worn-out hips, shoulders and knees. In typical anatomies, the choice between small, medium or large parts tends to solve most surgical-fit challenges. But congenital deformities tend to present as less symmetrical and uniquely shaped, requiring a highly customized approach.
A common birth defect provides an opportunity to heal thousands
One such deformity is pectus excavatum (funnel chest), a birth defect of the breastbone and ribs caused by an overgrowth of cartilage that skews the ribcage inward, resulting in an irregular, clearly visible crater in the upper torso. This malformation can occur in as many as 1 out of every 300 births, with men affected three to five times as often as women. Severe cases can impact heart and lung function, but thoracic-bone remodeling operations can be invasive and don’t always result in cardiac or respiratory improvement.
Milder cases of funnel chest are more common. While tolerable from a health point of view, they can nevertheless be cosmetically quite unusual-looking and cause a marked psychological impact. Correcting the anomaly must wait until someone is fully grown; the majority of patients are in the 18-35-year age range—and, by then, very eager to have an unremarkable torso.
Surgical attempts to correct pectus excavatum with metal inserts or springs date as far back as the early 1900s. But as those results proved less effective—with surgery deemed to be an excessive solution to cases that didn’t threaten internal organs—the use of implants to fill in the physical gaps above the bones became a less-invasive alternative. Prof. Chavoin’s work on silicone implants that could be inserted below the chest muscles but above the bones of the ribcage began in the 1990s. By the time Moreno connected with him, there was already significant market-demand for that solution.
CT-scan data analysis speeds implant design
Prof. Chavoin had been designing implants to correct pectus excavatum using plaster molds of individual patients’ chests—a rather time-consuming process that could only capture exterior anatomy. In 2008, he suggested to Moreno that CT scans might provide a fuller picture of patient anatomy on which to base more accurate implant designs. He was right: the CT-scan and computer-aided design technology delivered a heretofore unseen level of precision to the customization process.
Each device is uniquely fitted to the internal contours of a patient’s body, as well as external. Refined with the direct participation of the surgeon on each case, the new implants are manufactured to ISO 13485:2016 standards and delivered to the operating room more quickly for suturing inside the body in exactly the right orientation.
The solid implants, unlike fluid-filled ones, last a lifetime. For production, the AnatomikModeling team partnered with a certified medical device manufacturer that employs rapid prototyping to mold them out of polymerized silicone rubber from FDA-registered NuSil. The collaboration has proved highly successful in meeting market demand: they are now designing, manufacturing and delivering between 30 and 40 chest implants a month. The surgical technique they’ve developed has been deployed throughout Europe by a network of referral surgeons.
Tracking their success, the company reviewed some 20 years of patient files—starting from when plaster-mold procedures were used and then switching to the new CT-based methodology—and published a paper (based on random, blinded, independent review) comparing the two. Not only were malformations better corrected in the computer-aided design group, those patients reported higher satisfaction with the outcomes, both socially and emotionally. “The technique is simple and reliable and yields high-quality results,” the authors state.
A primary contributor to this success? Updated versions of the very same CT-scan data analysis software Moreno had begun using on bee brains and mummies.
Segmenting the data to pinpoint patient anatomy
Now, industrial CT scanning inspection of say, an automotive engine casing, is done with high-resolution X-rays that can penetrate almost any material. But because the human body must be scanned with as little radiation as possible, interpreting the data from a lower-resolution medical scan is a more nuanced capability: complex mathematical algorithms—in combination with user experience and expertise—are needed to process the imaging data into an accurate recreation of the patient’s anatomy.
“We’re essentially creating a digital twin of the patient so we can design the exact geometry of the medical device that solves the structural issues,” says Moreno.
“The challenge I’d first seen going from bees to mummies remains,” he says. “You need to scale up the methodology by segmenting the individual tissues. To digitally map a living human body so you can design an implant, you must identify and separate bone, muscle, cartilage, fat, and skin—each of which has a competing grey value.” His team uses Hexagon’s VGSTUDIO MAX software, with its Coordinate Measurement Module, to identify and isolate the segmentations that are regions of interest. From these, the 3D files of the different body tissues are calculated and exported to CAD modelling software for designing the final implant.
Beta testing AI – deep learning accelerates solutions
Early on, Moreno had been using the VG software Paint & Segment tool to isolate regions of interest in his scan data for further analysis. “Paint & Segment is a good tool, and it has been updated to be partially automated,” he says. Moreno’s company amassed a secure library of segmented patient data over many years. Then, in 2024, Hexagon proposed to AnatomikModeling that they test Beta versions of the new VGTRAINER application under development. “With VGTRAINER, we found we could automate all our segmentation steps,” says Moreno. “It’s extremely efficient and robust in terms of segmentation quality.”
Hexagon worked closely with Moreno and his team as they evaluated the new deep-learning software. “There is nothing more rewarding than collaborating and learning with such dedicated and skilled individuals,” says Hexagon’s Nicolas Coutant. “Benjamin started to use the tool by just following the tutorial and was really able to manage it by himself. He is not only a remarkable individual but also a passionate soul who approaches his work with great enthusiasm. He envisioned this AI workflow years ago; the tools were not available then, but the need was.”
Training the trainer
Deep Learning is a form of artificial intelligence (AI) that is increasingly useful for improving efficiencies across many fields, especially in the industrial realm. Because a manufacturer can keep its data firewalled and proprietary, industries such as aerospace, automotive, electronics and, in this case, medical are finding AI tools highly valuable.
“Training” is key to successfully employing deep learning tools, and training an accurate model is a critical first step that can take a while. But once the model is trained—in this case to identify individual tissue types from low-resolution scans—dramatic time savings are seen.
“When we began Beta testing, we started with 20 existing and accurately segmented patient data sets, fed them into VGTRAINER, ran it for about 36 hours, and ended up with a model that could segment subsequent datasets at about 98 per cent accuracy in just 15 to 20 seconds,” says Moreno. “We’re saving half an hour per patient case, time that can be better spent on added-value work; you are definitely getting a return on your investment in the software.”
How does AnatomikModeling confirm the accuracy of its deep-learning results? “VGTRAINER provides metrics that enable us to perform quality and robustness checks with our ISO 13485:2016 quality standards certification,” says Moreno. “In addition, since I’ve worked on more than 4,000 cases over 15 years, when I review segmentation results for each patient, I can visually, very quickly, see if we missed something or have incorrect anatomy. We always back up our work with human certification on the quality of results.”
Branching out to help patients with a wide variety of surgical needs
While some 90% of AnatomikModeling’s current work is focused on funnel-chest treatment, the company is collaborating with medical experts in other fields to develop implants for customized airway stents, jawbone reconstruction, skull-bone replacement, and atrophy of muscles in the limbs, as well as one-off applications for major injuries with unusual soft-tissue loss.
“Once you’ve trained your deep-learning model to identify all the different human tissues, it can understand muscle or bone, for instance, located anywhere in the body,” says Moreno. “We’ve now got the tools to treat a wide variety of implant cases. The software we’re using is top-notch in terms of stability, and you never experience any shutdowns or bugs because it’s made for the hardcore industry.
“Deep learning is speeding up our progress—though it always needs to be combined with human knowledge, the skill of a surgeon, and tracking patient results and satisfaction.”
