Artificial intelligence helps radiologists gain new insights from medical imaging. Angel Alberich-Bayarri about AI, biomarkers, and quantitative image analysis for the next generation of healthcare.
Wondering where AI is going and how your company can make the most of the AI revolution? Clusterone can help you gain strategic insights to get your company ready for the age of AI. Book a strategy session with us today!
When X-ray imaging first emerged in the late 1800s, a new era of medical diagnostics began. In the 20th century, other imaging techniques such as CT scans and MRI were developed and medical imaging has become one of the pillars of modern medicine.
Today, AI-aided image analysis has the potential to transform the field like once the development of the X-ray machine. While images quality has been steadily improving since Röntgen’s days, there is lots of headroom when it comes to the analysis of images.
A few companies around the world are developing AI-aided image analysis for the healthcare sector, but Angel Alberich-Bayarri and his company Quibim stand out. While others are still in the development stages, Quibim’s products have been actively used by hospitals and research projects for several years.
Identifying the need for quantification
During his Ph.D. in biomedical engineering, Angel did research on algorithms for post-processing of MRI images. When working at a hospital in Valencia, Spain after finishing his degree, he realized that many radiologists were looking for ways to quantify what they saw on the images.
Although a brain MRI clearly contained all the necessary information, there was no easy way of telling what the grey-matter volume of the brain was, for instance. His post-processing algorithms could do that job, Angel realized. And not only that: “I noticed that there were a lot of findings in the images that should be quantified, but they weren’t,” Angel says.
Angel’s ideas quickly became popular beyond his workplace and soon he found himself giving lectures about quantifying information in medical images to radiologists and other medical personnel. Clearly, there was a demand for such technology. Together with Prof. Luis Martí-Bonmatí, a leading radiologist in Spain, and other colleagues, Angel founded Quibim, a company to offer quantitative analysis of medical images.
Quibim helps doctors to do what they’re good at
Quibim automatically detects and measures biomarkers in medical images. They tap into the image repository of the hospital server and automatically analyze every image that comes in. Angel explains the process: “In the biggest hospital in Spain, we [analyze all relevant images] every night. Every night we pull all the cases, automatically analyze them, and send the results back to the hospital server.” The idea is, he says, that the analysis results are available to the physician as soon as they open the case file. If an analysis is needed immediately, it can also be triggered manually.
This means that instead of spending long hours on quantifying findings in images, physicians can do what they’re good at: interpreting the results and make decisions based on them.
Analysis results can be presented in different ways. “One [option] is that images can be enhanced with an overlay parameter, such as the cellularity in mm2/second,” Angel says. The software then creates a cellularity map that can be displayed together with the original image.
Another option is to produce a value sheet with relevant parameters, similar to what blood test results look like today.
Using deep learning for liver segmentation
Quibim uses deep learning and artificial intelligence for its analyses. “We love to use AI,” Angel says, “but we’re not an AI company.” For him, AI is a tool that simply is the best fit for Quibim’s specific needs.
Angel gives the example of measuring the fat content of a liver. To be able to automatically calculate this value, the analysis algorithm first needs to know what part of the image represents the liver. This process is called segmentation. “We use CNNs to create an automated way of liver segmentation. We have livers that have been segmented by an expert. Then, we develop a CNN network and train it with the human segmentations.”
To train a network to perform such segmentations, a large amount of training data is required. Quibim uses different ways of data augmentation to create large datasets based on relatively few liver images. “We can obtain good results with a low number of cases,” Angel says. Rotating or deforming the image and adding noise are only some of the techniques Quibim uses to augment its data.
Analyzing 50 million signals per second with neural networks
The company also uses neural networks to speed up calculations in their algorithms. The mathematical models Quibim uses to calculate cellularity or other parameters require lots of performance and time to compute when implemented in a traditional way. A well-trained neural network, on the other hand, can quickly solve the task using only a fraction of the resources. “An entire breast-MRI examination today has 50 million signals,” Angel says. “[Using a neural network], we can analyze these 50 million signals in 1 second.”
While he doesn’t want to call Quibim an AI company, Angel is convinced that deep learning and AI are the tools of the future when dealing with large amounts of data: “There is no way back. AI is the solution for the problems we had with traditional techniques in image analysis.”
Use AI to fight information loss
This is especially true when looking at how much data is produced by imaging technologies today. The data is so rich that it already is very difficult to be analyzed by human radiologists alone. It is very common, Angel explains, that radiologists reduce the amount of information in a scan to be able to handle the data. “A new-generation CT in a hospital [may have] a slice-thickness of 0.6mm,” he says. “Since that will generate 1000 slices, the radiologist says ‘let’s reconstruct at 2mm’.” This reduces the number of slices, but it also removes information that could be useful.
An AI-based algorithm can pre-process all slices and mark relevant information so that the radiologist only has to look at interesting areas. In that regard, companies like Quibim not only simplify work for radiologists, they actually enable a radiologist to make use of current generation scanning technology at all.
MRIs for routine checkups?
Angel’s goal is to become one of the leaders in providing medical image analysis worldwide. But his vision goes beyond the company. “Imaging examinations are at the end of the healthcare process today,” he explains. “Unless you have a symptom, you don’t get a scan.” In Angel’s opinion, this is too late in the process. Instead, he envisions non-ionizing imaging, such as ultrasound and MRIs, to become part of preventive medicine. “Every person should have an ultrasound or MRI scan [in certain intervals]. Then, we would be able to catch many diseases early,” he says.
Angel admits that these scans would be costly. But he is sure that in the long run, the savings would largely outnumber the additional costs. Ongoing studies in Spain and the Netherlands will show, he claims, that through such examinations, many diseases may be treated at a lower cost than if they worsen and become a life-threatening problem later on.
AI-based algorithms such as the ones Quibim is developing can help to analyze these routine scans faster and more precisely, even further reducing costs for scans. But of course, it’s not all about money. Regular ultrasound or MRI scans can also simply bring peace of mind.
“I have kids,” Angel explains. “I find it crazy that [they] grow up without knowing if they’re okay inside or not. How are the organs, is there a small tumor somewhere that will grow in the coming years? Can we cure it before symptoms [appear]?” With his work, Angel and many others in his industry will not only transform healthcare to be more efficient, save more lives, and reduce costs. With the help of AI, they may also help parents to sleep better, knowing themselves and their family are healthy.