Integrating artificial intelligence (AI) into the radiologist’s workflow is still challenging, but all-inclusive solutions can help unfold algorithms’ power, QUIBIM CEO and founder Angel Alberich-Bayarri explained during the ESR AI Premium Event, which was hosted by the European Society of Radiology (ESR) and the European School of Radiology (ESOR) earlier this month in Barcelona.
Medical imaging AI is a bubbling field but very few solutions are used in daily practice today, Alberich told delegates during the busy meeting, which gathered top researchers in medical imaging AI and thousands of online attendees. “We have a lot of research, AI algorithms and start-ups, but few are really embedded in the radiology workflow,” he said.
A main obstacle to AI integration is a lack of knowledge of utilities. For most imaging biomarkers, the real application relationship with clinical endpoints on a large scale – diagnostic, prognostic and treatment response – remains unknown. Clinicians don’t want to integrate biomarkers that have not been validated, but if they don’t gather information massively and try to understand how biomarkers relate to the disease, they will never help advance healthcare, Alberich explained.
“We have to change our minds and not wait for biomarkers to be validated before they can be extracted on a daily basis. Similarly to genomics: we have to do sequencing and study diseases to detect unknown mutations by ourselves,” he said.
The lack of annotated data is another challenge. Radiology reports are not filled with a focus on data annotation, but on descriptive language that needs to be processed by natural language processing (NLP). However they would be an ideal source of knowledge, and not just for the clinicians.
One-stop-shop platform using biomarkers
QUIBIM was created in 2015 as a spinoff company of La Fe Hospital in Valencia, Spain, to help radiologists make the most of AI in their everyday practice, by providing a one-stop-shop solution using imaging biomarkers and powerful algorithms.
“Offering a single solution embedded in workflows is key because radiologists will not buy all start-up micro developments but the best platform,” he said.
The very name of the company is an acronym and stands for Quantitative Imaging Biomarkers in Medicine. Imaging biomarkers enable to measure everything happening in the body to extract parameters that can provide information on the tissue of the lesion type beyond classification. Fuelled by AI, these biomarkers can deliver unprecedented information on disease.
“Many different imaging pipelines have dramatically changed thanks to AI integration. For example, we used to have lots of problems doing segmentation with traditional algorithms in tissues and organs such as the liver in MRI. Thanks to AI, segmentation has improved and with it our knowledge of liver disease,” he said.
A prerequisite is to integrate data mining solutions to make radiomics easy to everyone. This means it must be embedded in the same platform, as current solutions are not able to treat this amount of quantitative data from patient cohorts. “A lot of parameters are quantified in daily routine, but there is still no way to store and process them massively. Our current PACS systems are simply not prepared for quantitative data,” Alberich said.
There is a lot of sense in working within a structured report (SR), as it enables to annotate data that will further advance research. Prospectively the studies can be very well annotated if AI imaging biomarkers are integrated in the fields of the SR. Working with the SR would tremendously facilitate communication between clinicians and with patients.
“We are building the radiology report of the future, only we’re doing it now. Patients do not understand radiology reports, so we have to change the way we communicate. We are very much aligned with the standardized way blood test findings are reported, everyone understands whether findings are in or out of range and if there is any abnormality. We have to be more intuitive in our communication,” Alberich said.
QUIBIM is developing quantitative, one-page long structured reports that are actionable, quantitative and automated. The reports are designed with KOLs in each speciality, to make sure they reflect the reality of clinical practice.
Clinical input and powerful technology
Cooperation with clinicians is a key axis for the company, which uses a stepwise model to create new solutions with clinicians in the loop.
The company develops highly performing AI algorithms using unprecedented network architectures, for instance, multiscale convolutional networks ensemble in brain MR, 2.5D network in liver MR and referee network in chest x-ray, and hundreds of thousands of annotated images that have been acquired through cooperation with researchers and university hospitals.
QUIBIM has launched algorithms in different product areas to provide a head to toe imaging biomarker solution; in neuro for Alzheimer, ALS, MS, stroke and cerebrovascular events; in MSK for articulations; in chest for screening, COPD and fibrosis; and in breast for screening. More products will soon be released that will focus on other body areas.
More than 60 hospitals worldwide currently use QUIBIM tools, most of which have received CE marks and/or are FDA pending. The company has notably developed the new ESOR teaching platform and has supplied its QUIBIM Precision® image analysis platform to the AI Precision Health Institute at the University of Hawai‘i Cancer Center in Honolulu. QUIBIM has received €3.5m funding ever since its creation and has offices in Spain and the U.S.