[BELLE HISTOIRE] AI to optimize robot-assisted knee osteoarthritis surgery

Osteoarthritis of the knee is a common condition - affecting 30% of people aged 65 to 75, according to the French health insurance system- characterized by wear and tear of the knee's cartilage. As a result, the tibia and femur rub against each other in the joint, causing pain. If osteoarthritis becomes too disabling, arthroplasty is recommended: a prosthesis is fitted by an orthopedic surgeon, who has to cut away part of the bone to insert the artificial joint.

In such an operation, it is essential to know the characteristics of the patient's knee, in order to make the most accurate cuts possible. Traditionally, surgeons have relied on 2D X-rays, metal rods to be inserted into the tibia and femur, and their own knowledge and experience. " This method is, however, highly invasive, which can induce significant recovery time or risks of infection," observes Anna Gounot, PhD student at Ganymed Robotics. " Moreover, conventional arthroplasty has a high rate of patient dissatisfaction - 15-20% in the scientific literature. What ' s more, it is highly dependent on the practitioner's level of expertise. "

From scanner to robot-assisted knee surgery

Hence the interest in standardizing the intervention procedure, by assisting the surgeon with a collaborative robot. This is exactly what Ganymed Robotics is developing. Unlike other robotic solutions on the market, the French start-up's robot features a camera, which "sees" the tissue directly during the operation.

Like all robotic orthopedic surgery, Ganymed Robotics' approach begins with preoperative planning. First, the patient undergoes a CT scan to reconstruct an image of the tibia and femur, which is used to plan the bone cuts. Then, during the operation, this information must be provided to the robot so that it can position itself precisely in the planned cutting zones, thus facilitating the surgeon's work.

But it's not that simple. For example, while a CT scanner can show bones, it cannot display the cartilage within the knee. The presence of even this deteriorated soft tissue can have a major impact on such a meticulous operation. Consequently, the success of the operation depends heavily on the ability to predict its shape and thickness at the cutting points.

AI trained from MRI images

This is the aim of Anna Gounot's CIFRE thesis, the fruit of a partnership between Ganymed Robotics and the Laboratoire de traitement de l'information médicale (LaTIM, INSERM UMR 1101), under the joint supervision of IMT Atlantique, a component school of the Carnot TSN institute. Starting in April 2023, it is supervised by Valérie Burdin and Guillaume Dardenne on the academic side, and Marion Decrouez on the industrial side. The aim is to develop a method for automatically analyzing CT images to predict the distribution of cartilage within an individual's knee.

To achieve this, Anna Gounot has turned to a particular type of AI, belonging to the field of machine learning: the Statistical Shape Model (SSM). With this approach, the first step is to train the algorithm on a dataset to learn statistically the shape of a cartilage as a function of configurations. For this purpose, the research team relied on MRI images, an examination that shows both bone and cartilage. " Why bother with CT scans when MRI seems to provide more information, " says Anna Gounot. " But in reality, in clinical practice, MRI has its drawbacks, such as cost, certain contraindications, and less precise bone visualization than CT scans. Nevertheless, in research, it is possible to use specific MRI protocols that offer more precise rendering, including for bones. "

This is why the researchers have employed a database compiling a large quantity of MRI images of knees. " However, to exploit this database, a key segmentation step is required," explains Anna Gounot. " This means that on each MRI slice, we have to color the parts corresponding to bone and those where cartilage is present. A tedious manual task, but one that can also be automated. " As part of the thesis, the research team collaborated with other laboratories to obtain manual segmentations on MRI scans.

Predicting cartilage in relation to bone

For example, the researchers had the material they needed to drive the SSM developed by Anna Gounot. Or, rather, SSMs. " We decided to develop two different SSMs," says the doctoral student. " One was trained solely on healthy individuals, the other on patients suffering from medial osteoarthritis of the knee, i.e. affecting the medial part of the joint. While the former was not of direct interest to Ganymed Robotics, it did help us to better understand how an SSM works, and to observe potential learning disparities from two different databases. "

However, Anna Gounot wanted to enrich this training phase by visiting Professor Bhushan Borotikar of the Symbiosis Centre for Medical Image Analysis (SCMIA) in Pune, India. He is conducting a campaign to collect data from cadavers, by having them undergo scans and MRIs. " This enabled us to train our models on different data, which can only improve them," says the doctoral student. " And we also had access to CT images, not just MRI ones. This brought us closer to the real-life use case for Ganymed Robotics, since we'll be dealing with patients undergoing CT scans. Finally, this three-month stay was an opportunity for me to learn manual segmentation and develop a deep learning model to automate this task. "

From these diverse data sources, SSM have learned to establish a statistical link between the shape of a tibia and femur and the associated cartilage. A skill then used to automatically analyze a scanner image and provide a map of the likely distribution of cartilage on a bone.

From statistical cartilage prediction to patient-specific prediction

"It' s important to remember that, at the moment, we're only talking about statistical predictions, based on SSM learning and scanner images, in which cartilage does not appear ", insists Anna Gounot. Yet osteoarthritis remains a complex pathology, with each patient presenting marked singularities. Consequently, any specific information is invaluable for improving and individualizing the model's predictions.

To this end, the Ganymed Robotics robot is equipped with a 3D camera providing a 3D reconstruction of the joint in the form of a point cloud. This information is particularly useful for the model. As this robot is still under development, the start-up collected data by placing 3D cameras during real surgical operations. The images collected could be used by the research team, along with a synthetic database generated by the researchers from existing algorithms.

For example, SSM has improved the accuracy of its cartilage predictions, making them more patient-specific. These are results that Anna Gounot will be looking to develop further, including at the end of her thesis, when she is hired by Ganymed Robotics. And although they currently only concern knee arthroplasty, they could be transposed to other orthopaedic surgeries, notably shoulder or hip surgery.