Paris Innovation Review - Medical imaging is one of the most active fields in terms of research and innovation. Where does the technological frontier stand today?
Wafa Skalli - This is a very broad domain, covering very different disciplines: advances are numerous and at once concern hardware, interpretation techniques or clinical applications. While we won’t be reviewing the whole gamut in this paper, a few particularly dynamic areas can be pointed out. I would eagerly cite neuroimaging (techniques allowing to observe the brain), which is opening perspectives on some items that we did not expect, not only regarding organic diseases but also about the way the human brain and thoughts actually function.
Another booming area is ultrasound elastography, which allows to measure tissue elasticity: it is a key issue for the clinical diagnosis of certain diseases, cancers in particular. Finally, there is extensive research on the skeleton, which is now enabling significant advances in our understanding of aging mechanisms, or of other pathologies such as scoliosis.
Let us take the case of 3D modeling of the skeleton that biomechanics lab researchers are working on. Where is innovation playing out the most: on the hardware side, or rather, on the side of methods in reading and interpreting of images?
Both areas are complementary and their progress is parallel. In terms of techniques and equipment, the main challenge is to gain precision and get to see as much as possible, and in the least invasive, least irradiating manner for the patient.
The Eos system was developed with this in mind. It allows to reconstruct a three-dimensional image of the spine out of only two x-rays – and low-dose x-rays at that, which are a priori less accurate than a scanner. But from these two radiographs, we obtain a 3D representation which would require 300 cuts if we were to use a scanner. Much of the innovation effort here resides in the software dealing with 3D reconstruction.
Processing medical images is therefore a key issue: the aim is indeed to automate much of this process. The construction of the 3D image is in itself an interpretation, by the machine, of the raw data obtained from the two x-rays. While the resulting images are indeed rich with information, they are also parasitized by “noise”. A clinician can tell the difference right away: for example ribs appear as regular bands. But such knowledge, which comes from experience, is not easy to automate. How can we teach a machine to identify human ribs? This is one of the challenges of our work.
This requires a multidisciplinary approach which one can easily imagine is far from obvious...
Yes, there is still strong compartmentalization between the disciplines, but disruptive innovation is precisely to be found at the crossroads of different fields. Eos resulted from the encounter between the French Nobel prize physicist Georges Charpak and two professors of medicine, Jérôme Kalifa and Jean Dubousset. In the world of traditional research, this event would not have occurred, but in this particular case a project sprung up very quickly. The Biomecam academic chair of Arts et Métiers ParisTech (an engineering school) perpetuates the legacy of this multidisciplinary approach, bringing together researchers from different specialties (from physics to biology through osteoarticular radiology), who get to work with clinicians, software developers... What is at stake is therefore not only to get disciplines to converge and intersect, but professional fields as well.
Moreover, this is the very concern we had in mind when we injected a strong interdisciplinary approach to our curriculum for the ParisTech Medical Engineering Master, which aims to train professionals who, despite being specialists in a given field, also possess a good knowledge of what is at play elsewhere. Students revel in this, and the overall light they can cast on the problems of medical imaging is an extremely valuable resource in the practice of their future profession.
You mentioned a fertile convergence of different professional fields. Doesn’t this also apply to a link between research and industrial applications?
This is obviously essential. Our point is not just to advance our work in order to fuel further research, but to raise awareness about the socio-economic benefits of innovation. This means being mindful of the protection and advancement of intellectual work, being able to move from innovation to entrepreneurship, but also and more broadly being able to assess the impact of our work on clinical practice and therapeutic choices.
Let me provide you with an example: doctors have long been able to identify scoliosis in children aged eleven or twelve, but it is very difficult, at an early stage, to know whether a slight deformation will worsen or not. The stake is an extremely sensitive issue for the patient: will he or she be forced to wear a medical brace to prevent future disability? There is a risk of making that choice when it is actually not necessary, just as it is also possible to decide to do nothing, while in fact dealing with a heavy pathology. Empowering the physician with new tools to refine his diagnosis is decisive for the fate of the young patient, but also constitutes a key economic issue for health systems.
The Eos system that you use for the 3D modeling of the human skeleton also happens to be the result of collaboration between a business and public institutions.
Yes, and its name was chosen by the Biospace Instruments company, created by Georges Charpak himself. This is not an acronym, as is generally believed; it is the name of a Greek goddess – the goddess of the dawn. A name that also evokes the bones in French (os means bone), as it revolves around our skeleton.
Eos is chiefly based on two elements, and both are disruptive innovations – one is based on research conducted by Biospace Instruments, the other on public research.
First innovation: a special technique of digital radiography, thanks to which Georges Charpak won the Nobel Prize in Physics. This technique uses high-sensitivity detectors, the “multiwire proportional chamber”: containers filled with a noble gas, argon, within which parallel grids are installed that are made up of many wires. This device is capable of detecting ionized particles with extreme sensitivity and accuracy. This allows to obtain a very good image quality – even with a very low irradiation.
The second innovation is to generate not just one single x-ray, but a pair – front and profile – all with one sweep of about twenty seconds. From these two perpendicular planes, a software allows to reconstruct an image of the skeleton in three dimensions.
These two innovations allow for progresses that are also disruptive innovations. For instance, to obtain 3D with a scanner or magnetic resonance imaging (MRI), a patient must be installed supine, that is, lying on his back ; with Eos, he or she is standing – which is especially useful when one is dealing with diseases located on the spine, but also on the hip, knee, the foot: if you wish to understand the pathology, better have a patient standing.
Yet another notable achievement: we have overall vision, whereas CT scanning and MRI are limited to a particular region. With Eos, one can render the entire spine, or leg – and a few years from now, we will be able to render the entire body. This broadening of perspective is essential: it allows digital rendering to scale up, matching the scale of the human body and not just working at the scale of single organs. This is extremely useful and relates to what doctors have known all along: the deformation of a joint is always associated with other disorders. For example a deformed knee will be matched with a certain stance that will keep impacting the spine, the ankle, etc. These are complex relationships, whose understanding is in no way obvious: the compensation phenomena need to be measured, quantified, and modeled. The causal order should, when possible, be reconstituted...
One can easily imagine that the scope this is opening for research is huge, but how will research be organized? For instance does it proceed one pathology or one organ at a time, does it focus more on understanding a particular case? Or is it quite the opposite: do you resort to comparisons, or even to statistics?
It should be remembered that one of our challenges is to improve Eos. One of the directions of research today is to work on three-dimensional reconstruction techniques: to make them more accurate, more automated. What is more, a number of researchers are already working using Eos. For instance, to take an example mentioned earlier, to understand why some scoliosis cases get worse and not others. Another subject, related to aging: to understand the underlying relationship between pelvis, spine and head posture.
Most of the research around Eos is co-led by an engineer and a clinician. The goal here is to articulate the improvement of techniques with the needs expressed by doctors in terms of diagnosis and comprehension of diseases.
But I would not want to give the impression that we only conduct our work around human body malfunctions. Clinical innovation cannot be conceived of without basic research. It is essential, in medicine, to understand “normality”, that is to say the normal functioning of organs in the absence of pathology.
Let us consider something that is both a very common experience to everybody, and a very complex system: walking. Questions that can arise around this issue are very diverse. One of these questions is how the neuromotor control system mobilizes sensors to regulate and control an erect posture and normal walking.
A better understanding of what happens when everything is in order helps answer some very specific clinical questions. For example, a child with cerebral palsy cannot walk normally and there is muscular hypercontraction. Not only is the child disabled, but his bones will deform, too. How can we improve these children’s condition and prevent this distortion? We are faced with a problem of therapeutic choice: we can work on relaxing these muscles, or decide to move them. To help the medical profession make these choices, it is essential to better understand the very complex functioning of the human body and what functions are mobilized when we “simply” walk.
One gets the feeling, listening to you, that while the share of applied research (clinical, in this case) and basic research must be determined, back and forth exchanges are constant between the two areas.
They certainly do feed each other. This is particularly noticeable if we take other directions of research. For instance, automobile security in the case of car accidents. We have long collaborated with car manufacturers, and there is also a joint laboratory of biomechanics and accidentology involving Renault and PSA. One of the challenges of innovation in the automotive field is to protect the occupant during an impact. Our research interests them, and conversely their knowledge in accidentology enriches our representation and our understanding of how the human body functions.
The same is true in another area that is highly instructive and which is associated with specific pathologies: sport. Or, even if research is still too rare in this area, occupational diseases and in particular musculoskeletal disorders that appear repeatedly in association with a given activity.
The latter example opens an epidemiological and statistical approach, where the images produced by Eos also have a part to play. Recently, an idea has emerged: research, besides the deductive model that has dominated so far, could also operate though inductive logic – for example with the automatic comparison of thousands of x-rays.
In this respect the lines have indeed been shifting, owing to the ability to exploit massive amounts of data to extract knowledge. There are now two chief models: knowledge deducted from physical observation, and statistical models derived from data mining, where laws and trends appear without having been previously imagined or observed directly.
This is doubtlessly a profound change in the very way that scientific knowledge is generated, in the biological field at any rate. However I shall point out that in order to be effective these processes must still be guided by an expert’s knowledge.
Thus, image processing is not completely automated; one must therefore ensure that it is as robust and as fast as possible so as to optimize the share of interpretation that ends in the hands of the expert. Practically, this means for example producing images that offer maximum clarity to the human eye. But the human eye is not that of the machine, and there are trade-offs involved here, workarounds have to be articulated.
Whether we’re dealing with a machine or a human being, an important issue is that of uncertainty, which involves the issue of validation. Especially for the machine, it can translate into a quality approach. The idea is both to ensure the reliability of automated readings, to set standards, and to quantify the amount of uncertainty.
You see, the new tools and the methods allowing to exploit them on a large scale do not, cannot replace the expert. The latter simply sees his or her role shift.
New issues appear at this stage. Some have to do with collection protocols. Let us come back to the case of Eos. There are currently twenty centers using this technology. If you want to exploit their images in an automated frame, then you have a requisite that protocols have to be compatible.
Ethical issues arise from some of the methodological issues involved. We handle medical data that involves people: it should be ensured that they are anonymous, yet without screening out any information that might be useful to the researcher, such as age, sex, etc.
Does the tension that one perceives between the “two chief models” that you mentioned earlier lead to dilemmas, such as the painful necessity to choose between the path of the machine or a support to the expert?
I would more readily call this a productive tension, and furthermore, more than a tension between the logic of the machine and human expertise, a tension between clinical science and epidemiology, between the peculiarities of a given individual and the general laws governing the human body at work.
Such a tension is productive. I would like to explain it by mentioning a recent example: knee replacements. We now know that there is a 15% complication rate in ten years’ time – a considerable challenge if we take into account what this may personally translate into for patients, but also in terms of costs to society. Here we are right at the intersection between individual cases and a statistic. This 15% figure represents something, and at the same time each patient is unique. How can we anticipate the risk of complications? The solution that is emerging today comes in the form of custom models – that is to say the combination of methods derived from general laws, and an in-depth consideration of the individual. What we end up with is close to the “hunch” or “gut feeling” approach of a good surgeon; except for the fact that models add some very powerful objective elements to the equation.
That goal – succeeding in modeling individuals – constitutes a new technological frontier, but at a deeper level it is also a renewal of scientific discovery itself. What we are experiencing, both in the knowledge generated and in the way to generate it, is nothing short of a quantum leap.