You’ve probably heard about our revolutionary product called BoneMRI, and maybe you’ve been wondering about all the “magic” happening behind the scenes. In this blog post we explain how BoneMRI uses Artificial Intelligence (AI) to create synthetic CT images from MRI.
When you think of AI you probably think of your nightly conversations with ChatGPT, your artwork created by DALL·E2 and all the random faces generated by This Person Does Not Exist. Let’s clarify, these are wonderful tools based on AI, but this is absolutely NOT comparable to the BoneMRI algorithm. It would be foolish if we would just generate random CT images to diagnose an actual patient. Wouldn’t be very safe either!
CT scan generated by DALL·E2: This is not what we do….
MRI versus CT
First, let’s have a look at how it all started. For the newbies in medical imaging, CT uses X-rays with ionizing radiation to image the inside of your body. On the other hand, MRI uses a powerful but harmless magnet and radiofrequency waves to do the same. In general CT is better for visualizing bone, and MRI better for softer tissues like the brain, organs or nerves. For specific indications one modality might be preferred above the other, but overall CT is cheaper and faster. So why use MRI instead? Because CT radiation exposes patients, especially children, to increased risk of cancer.
So what if… we could create an algorithm that transforms MRI to CT-like images? It was with these thoughts that Peter Seevinck started MRIguidance 6 years ago.
MRI, BoneMRI and CT of the lumbar spine
Although MRI and CT visualize the same structures, when you compare both modalities the image intensities are completely different. On MRI many structures that are coloured black are bright white on CT. Yes, we’re talking bones here. Not too strange that the first BoneMRI algorithm was based on taking the inverse of an MRI. If you observe the resulting image you can already guess that it takes more than that. We need something better. We need deep learning.
First ever BoneMRI without deep learning vs. BoneMRI today
Where the magic happens
Deep learning is a form of AI where you “train” a model based on an input to create a desired output. The classic example: you show the model many pictures of cats and dogs, and every time it incorrectly tells you what it’s looking at, you correct it. This way the model “learns” by optimizing its parameters until reaching a score you’re happy with.
The same trick can be applied for many different goals. In the case of BoneMRI, we use deep learning for image synthesis. This means that based on an input (MRI image) the deep learning model creates an output image (CT image). To train a BoneMRI model we present multiple MRI scans as input and use a CT scan as desired output.
After many iterations, the model learns to transform an MRI voxel (= 3D pixel) to the intensity value that would be expected for a CT at that location. This is already where we distinguish ourselves from the popular AI tools mentioned before. While these tools allow a lot of freedom in their output, the training of a BoneMRI model happens under very strict constraints with regard to the output. During model training the predicted BoneMRI output is always matched to an actual CT scan of the same patient, anatomically in a voxel-to-voxel approach.
Deep learning model for training BoneMRI
Garbage in, garbage out
A popular motto in our deep learning team is “before you start cooking, you need good ingredients”. This is not because we all like spending hours in the kitchen… It is because you can build the most beautiful deep learning architectures, but if the training data is of bad quality, you will never achieve good results.
For BoneMRI this process already starts at the MRI scanner. One of the reasons why we can generate accurate high resolution 3D BoneMRI images with our product is because we fully control the MRI sequence in the hospital. We are not dependent on standard clinical MRI scans, which are often in 2D and are never the same throughout hospitals. We define our own MRI sequence that acquires 3D submillimeter high resolution images.
Having a custom MRI sequence is already a good start in standardizing our data, which is an important concept in deep learning. But for BoneMRI training we are also dependent on the CT to which we optimize the model. To do this in a voxel-to-voxel manner, we need the CT to match the MRI exactly. Therefore, MRI and CT data are acquired from the same subject. However, there is time in between both scans and the subject might be scanned in a different position. As a result the MRI and CT data we receive are not aligned. This makes image registration an important data processing step to prepare the data for model training.
With image registration we align each voxel on the MRI to a voxel on the CT that matches the same position in the same anatomical structure. By carefully executing this process, we ensure that the BoneMRI model is trained using an input MRI that is exactly matched to the desired output (reference CT).
MRI, BoneMRI and registered CT of a vertebrae all perfectly align
Last but not least, in deep learning it is important that a model is robust when encountering real-life situations. This is extra important when your device operates in the medical world (like ours).
During training, the deep learning model should be exposed to as many data variations as possible. In our case, this means for example MRI scans of different vendors, magnetic field strengths, anatomies and patient populations. To obtain such data we proactively conduct extensive clinical studies and have collaborations with many hospitals from all over the world. This allows us to train our models on a broad range of examples reflecting what can be expected in clinical practice, and importantly, avoiding biases.
Metrics are key
BoneMRI technology is clinically validated by performing independent studies and extensive evaluations before we apply it to actual patients. This is important to ensure that the quality of BoneMRI will remain high when used in a new hospital. During development many metrics are used to monitor model performance. Together with qualitative reviews by clinical experts, these form the foundation of our validation process.
An example of such a metric is the deviation in bone morphology on the BoneMRI versus CT. We make sure that the bone surface on the BoneMRI will not deviate more than 1mm from what you could see on a CT. Because medical images are used for surgical planning and navigation, a submillimeter accuracy is of high importance. We don’t want a screw implant ending up in the wrong place. This is only one of many applications for BoneMRI, but it shows why the validation process plays an important part of the work done by our deep learning team.
Analysis of the morphological accuracy of a bone on BoneMRI compared to CT (left), important for the planning of screw implants (right)
The BoneMRI we generate with our deep learning solution is supported by clinical studies, CE and FDA clearances, and hospitals that use it in daily clinical practice. However, the world of deep learning never sleeps and anticipating the future is important to move along with the needs in the medical world. Therefore, we keep improving our technology and we keep challenging ourselves to use the most state-of-the-art techniques.
By understanding the basic concepts of deep learning, you’ve gained insight into why BoneMRI is a reliable, high-end medical product. Our blog will continue to explore the world of AI and deep learning. We will also cover topics related to clinical validation and regulatory clearances. You can find these blog posts in our quarterly newsletter, which you can subscribe to here. You can also follow us on social media for updates: check us out on LinkedIn and YouTube.