Artificial intelligence is being used more and more often in the healthcare sector too. AI can help detect certain illness profiles earlier and more quickly on scans. But because new upgrades often become available, the proper operation of AI-based medical equipment always needs to be verified and validated again. Through the European Vivaldy project, along with Barco and icometrix, VITO has developed some solutions for making this process quicker and more cost-effective, without compromising on quality.
The Leuven-based company icometrix specialises in the development of software for medical scanning equipment, such as MRI and CT scanners. Their algorithms can quickly recognise and quantify tissue damage, including in the brain. Doctors send their patients' three-dimensional scans to the company for this, which then analyses them and sends the results back. ‘We were primarily working on multiple sclerosis, but have evolved towards a digital health tech company that's focused on other neuro-degenerative conditions like Alzheimer's and epilepsy, as well as brain trauma too,’ says Jan Verheyden, VP for Partnerships at icometrix.
Continual learning process
The analysis of the (brain) scans by icometrix is based on deep learning, an AI technique whereby algorithms are fed (‘trained’) with a huge number of reference scans, through which they learn, as it were, to recognise tissue damage on new scans. This learning process is never complete, however, as new ‘training data’ is constantly becoming available. Verheyden: ‘Such new data might come from scans of people from new geographical regions or population groups, for example. The advantage of our AI techniques is that this new data can be fed into them very fast, meaning the entire software can be upgraded quickly.’
Unfortunately, these rapid, continual upgrades are not presently leading to improvements to the AI used in medical practice at the same pace. This is because it involves medical equipment, which is subject to very strict regulations. Before such equipment comes to market and is allowed to be used, it must first be comprehensively evaluated, validated and certified by government agencies. This also applies to the medical AI software. ‘For every upgrade, we have to draw up and submit a new evaluation case. So that takes a lot of administration, which is expensive and time-consuming.’
This was why, along with Barco, which also specialises in medical imaging technology (among other things), and VITO, icometrix joined the European Vivaldy project. In Vivaldy (which refers to ‘VerIfication and Validation of Ai-enabLeD sYstems’), specific solutions and applications were being sought for a quicker and more cost-efficient evaluation process. The Flemish branch of the project focused on several use cases, including counting and measuring lesions (tissue damage) on MRI scans on MRI scans (icometrix), and the identification of potential skin tumours on photos (Barco).
What was VITO's role in this project, which ran from early 2020 to early 2023? ‘We were the research partner in the collaboration,’ says Bart Elen from VITO. ‘With our AI4Health platform, we have a lot of experience with medical applications for artificial intelligence. This even ended up with the creation of a VITO / KU Leuven spin-off (MONA), which helps track down eye illnesses based on scans of the cornea.’
In Vivaldy, the work of the VITO experts was more fundamentally targeted. Elen: ‘Among other things, we're focusing on AI recognition of special cases, where the algorithms analysing the scans might have more difficulty. These include patients with tattoos on their skin when applied dermatologically, but also patients of a different ethnicity to the patients on whose data the algorithms were trained.’
Better performance and monitoring
The solutions that came about in the Vivaldy project involve better performance, which is guaranteed and also proven by the developers of the medical AI. ‘Through performance metrics, we can proactively convince regulatory bodies, which saves time. In addition, upgrades are being better monitored by human specialists as well. If the AI makes an error, they can intervene.’
The solutions have already had a good reception, both with regulatory bodies in Flanders and Europe, and in the United States. ‘The visit to the FDA (the competent inspection authority in the US) was a high point for us,’ says Verheyden. ‘Normally, these kinds of bodies are relatively tight-lipped. Thanks to this project, we got the chance to see how things work in practice from their side as well.’
Barco is already applying the solutions in its own R&D division. ‘For example, we're using them to identify trends for new, upgraded algorithms, which we can then take with us to the evaluation,’ says Tom Kimpe, from Barco's healthcare division. ‘That means it goes more smoothly and the validation and certification come more quickly.’
The Flemish branch of Vivaldy was supported by VLAIO, which sees a strong strategic use in the project. ‘The results of this project, in the form of the solutions developed, are generic and thus widely applicable,’ says Jonas Van de Vyver, AI and ICT project advisor at VLAIO. ‘And in fact, it makes periodic retraining and further training of AI more important in many sectors – not just in healthcare. Facilitating that is important to Flanders.’
With the new experience it gained in the project, VITO itself is aiming to continue on its present course of more fundamental AI research. It will also do so in wider European collaborations, such as REALM (an initiative for basing healthcare policy more on real data) and TEF-Health (a testing and experimentation environment for medical AI and robotics).
This research was funded by the Vivaldy project, PENTA 19021, and financially supported by the Flemish Government HBC.2019.2714.
