Ángel Alberich-Bayarri, PhD, has long been fascinated by the fact that the human body generates signals and the ways those signals can be captured and reconstructed. As an electronics and computer science engineer interested in computer vision, he began investigating how various machines use physics to capture human body images.
“MRI, for example, is a modality in which you energize the tissue, and the tissue is giving you back a signal that you can capture and build an image,” Alberich-Bayarri told Inside Precision Medicine. “As an engineer, I was very excited about how complex the human body is and how you can use physics to generate those signals. A CT scan uses an X-ray, and [any signal] the body attenuates doesn’t go to the detector. The same goes for PET—you inject a radiotracer that sends the signals. The beauty is that you can capture all those components and build imaging data.”
After finishing his graduate studies, Alberich-Bayarri worked at a hospital to optimize image acquisition, where he became interested in quantifying various aspects of an image. He also began to notice some flaws in the use and application of medical imaging, such as a lack of algorithms within the various imaging devices in healthcare settings.
“I had family members who had a disease about which I would have published a paper, but when they went to the hospital, they did not have this kind of analysis,” said Alberich-Bayarri. “I saw that the original equipment manufacturers were not as quick to develop that, and we already had algorithms, so I said, ‘Okay, I can take that to market.’”
At 30, Alberich-Bayarri founded Quibim—short for “Quantitative Imaging Biomarkers In Medicine”—in his hometown of Valencia, Spain, to accelerate the integration of algorithms and advanced analytical technologies into clinical practice with AI-powered imaging biomarkers. The company’s first years were spent transitioning all Alberich-Bayarri’s research, code, and projects into regulated medical devices—a resource-heavy and costly endeavor. Notably, commercializing these types of models requires massive amounts of compute power to train the AI-based algorithms and must be validated in clinical studies before they can be regulated as medical devices.
Today, Alberich-Bayarri and Quibim received a boost as a healthtech company with the announcement of their $50 million Series A financing close. “We are using the proceeds of the funding round to differentiate ourselves as a category company in precision medicine, helping biopharma companies to stratify patients better,” said Alberich-Bayarri. “Ultimately, that could lead to new company diagnostics and new digital biomarkers that can be linked to a drug or not. Maybe they can explain a lung cancer patient’s response or overall survival. So that’s the main focus.”
A new category of precision medicine
Historically, biopharma images have been at the tissue or cellular level (or both). Patients receive an MRI or CT scan to obtain images of the brain. X-rays and CT scans may be used to diagnose lung cancer.
Pathology is one of the most significant areas where AI has been used in medicine, thanks to extensive research into image processing algorithms. With so much imaging data available, many institutions and companies have been racing to develop the most advanced foundation models, which are models developed without human supervision—that is, the data used to train the model has no labels indicating whether a particular image is from a healthy or diseased patient.
Alberich-Bayarri aims to take AI-based imaging biomarkers further to become a category company—businesses that define a new product category or service, such as Uber creating the ride-sharing world. Because biopharma companies have a wealth of imaging data accumulated over years and years of clinical trials, he believes there is an opportunity to use AI to create a new category of using images for personalized medicine, similar to what others have done in genomics, where it was not just about analyzing DNA. It spilled out like mutations and about researching and identifying new genomic panels. This data can be used to develop new diagnostics, such as selecting better patients for clinical trials or predicting drug response.
“You can decide to make a model simple and train it with what an expert would say—in this case, the radiologist—but if I have to sell this tool to radiologists, then the tool will be trained to identify things that an expert radiologist would say,” said Alberich-Bayarri. “That’s not adding value if the radiologists are not good enough—the algorithm will only be as good as them. If you want to add significant value, you need to train the models with something that’s in the biology, that’s not in the MRI, but AI can learn to link MRI to biology.”
Training AI with multimodal data
So, how is Quibim doing this? Instead of showing the AI a biopsy, Quibim provides other data modalities, such as survival rates, or an orthogonal imaging approach, such as MRI and CT scans. With this multimodal data compiled and analyzed by an unsupervised learning algorithm, the models learn to predict features of the biopsy or scans beyond whether a lesion exists by using these additional data modalities, which could be predictive of future outcomes or recommend treatment strategies.
Alberich-Bayarri said, “I’m not only informing you about this lesion here; I’m also telling you that based on my imaging data, this patient is at risk because the same patients that have these characteristics based on AI, we got to know now that they have a low survival, so you better take a more intense treatment on this patient because, based on imaging, this is a high-risk patient. So that’s the kind of information we want to create. We need money for computing. We need to train AI models. AI models apply to imaging.”
For example, Quibim’s QP Prostate, an AI-based solution for prostate MRI that automates prostate segmentation and lesion detection, has not been trained to detect cancer based on a radiologist’s image interpretation. Instead, Alberich-Bayarri and Quibim trained an AI model on EMR and biopsy data to identify MRI lesions that a radiologist may miss but are positive in a biopsy.
“When a radiologist is using our QP prostate, for example, it’s quite frequent to say, I see this lesion, but Quibim is saying it’s negative—there is no lesion there,” Alberich-Bayarri explained. “The algorithm has learned to identify lesions in a biopsy and match the patterns of the MRI. That multimodal approach to training the algorithms makes the radiologists perform better when they go to a tumor board. We don’t want to create a workflow improvement company for radiologists. We want to add value to the whole community, biopharma, the oncologist, and to the clinicians asking for a higher value of data from the images, more than just a report from the radiologist. If we can achieve that, we make radiologists like having a higher value on what they generate.”
Whole-body profiling
While the idea of using any image is appealing, in the real world, so many people use various instruments worldwide that it is difficult to imagine how one algorithm could be applied in a setting. Alberich-Bayarri has worked to create products that are compatible with a variety of instruments and platforms. To accomplish this, Quibim created a foundational model for harmonizing all the image qualities scattered worldwide—whether Siemens, Phillips, or GE machines in locations across the United States, Europe, and the United Kingdom.
“We have not seen many companies taking care of harmonization,” said Alberich-Bayarri. “We’ve seen companies going straight to putting images and trying to get the most out of it. But this is not Instagram. You don’t have millions of patients with pancreatic cancer. That means we have to work in conditions of low numbers of patients. We can synthesize data and create data augmentation, but you need powerful data harmonization techniques. Otherwise, you will not extract the most out of these images.”
If enough data exists from any imaging modality, Quibim could theoretically develop a solution for even more complex imaging types, such as whole-body scans. Alberich-Bayarri stated, “In the coming years, we will have more access to whole-body scans, which will allow AI to move from being fragmented in medical imaging to being something you can use across the entire body, right, with algorithms that can analyze all tissue parts, and the foundational algorithms will become better and better.”
“My long-term goal has always been to use imaging to characterize what’s happening at every tissue point of the human body,” said Alberich-Bayarri. “I believe that the genotype is information; you can do a genome analysis, but that genome might not precisely express in a certain way in the human body. So, I might have a predisposition for a disease, but I also might not have it. The only way you can understand the phenotype and what’s happening is by looking at your images from the inner body. What I want is to be able to use these images to characterize ourselves completely.”
