Investment in artificial intelligence (AI) development has slowed down and the community must look for alliances with partners with a high computational capacity and strong scientific background, a Spanish expert explained at the Triangle meeting last January in Madrid.
The development of AI has created a paradigm shift in research, Angel Alberich Bayarri, CEO of Quibim, told the audience at Triángulo, the annual gathering of Spanish radiology’s crème de la crème.
‘We are used to seeing medical research from an academic point of view,’ he told delegates. ‘But now, in the hospital, industry is leading research efforts.’
Today the most important frameworks used for AI development such as PyTorch (Meta) and TensorFlow (Google) have been made available by tech players. ‘Every AI researcher in academia, industry and research institutions in the world is working with these infrastructures,’ he said.
Following the Large Language Models’ trend, companies such as Google or OpenAI (Microsoft) that have a high computational capacity have taken the lead.
‘OpenAI is trained 24/7 by thousands of GPUs, over a large period of time,’ Alberich said. ‘That’s a cost that cannot be borne by almost any public institution yet.’
In Europe, advanced computational centers are capable of computing at most 4,500 GPUs over the same period of time. ‘We would need a CERN-type initiative that is clearly dedicated to AI in order to be able to provide such horsepower,’ he said. ‘We don’t have that right now.’
Selecting the right partner
In 2019, before the pandemic, specialists including Alberich had warned that the investment bubble in radiology AI would burst. Four years later, their predictions have come true.
‘Last year we’ve had the least investment in AI ever,’ he said. ‘There was a relevant jump in company closures in Q1 – Q2 2022. Since the U.S. Federal Reserve has increased interest rates, it is more challenging to move the same money from investors’ pockets to healthcare AI companies.’
The market has entered a phase of consolidation, and radiologists are selecting those products that are going to be more valuable in clinical practice.
‘Radiologists have had to spend a lot of time to understand these algorithms that were going to end up in hospitals,’ he said. ‘Looking for solutions has meant a great deal of effort.’
A recent publication in the Journal of the American College of Radiology (1) predicted that, by 2035, about 350 AI products with FDA approval would be available on the market.
Many companies will have disappeared then, and those that remain will either be giants ‘that have outlived the .com’ such as Google or Amazon, or those companies that generate a lot of scientific evidence to implement AI in hospitals, he believes.
A pharma of algorithms
Quibim has published its research model in Nature – Biopharma Dealmakers in June 2023. ‘We start by doing very strong research with pharmaceutical companies and academia, working with population data, and from there, advance product lines and obtain certification up until we reach patient level,’ he said. ‘It’s as if we were a pharma of algorithms. With one essential difference: the discovery of new imaging biomarkers is done organically by our team.’
The company’s QP-Prostate suite is already commercialized in Europe and the U.S, and it has recently received CE and UKCA approval for prostate cancer detection.
‘Our algorithm is trained with anatomic pathology, and that’s the big difference with other companies’ solutions,’ he said.
Quibim has recently partnered with Philips, and the software is now available directly inside the MR machines in the hospital.
The company also offers solutions for immunology and lung cancer through partnerships with Novartis, Merck and Bayer, as well as breast, colorectal, liver and brain lesions.
‘It’s important that we not only develop algorithms that detect lesions marketed in radiology as workflow improvement tools, but that we also go to predictive biomarkers, to be able to provide more information on diagnosis and treatment response,’ he said.
Quibim also works with patient groups, following the implementation of the new European health data space directive, which places the patient at the center of care.
(1): https://www.jacr.org/article/S1546-1440(23)00647-6/fulltext
