Cognita develops foundation models designed to interpret medical images end-to-end and draft comprehensive radiology results, initially focusing on X-rays and CT scans while keeping radiologists in control of the final report. Pulse 2.0 interviewed Cognita CTO Zhihong Chen to learn more.
Zhihong Chen’s Background
When asked about his background and the experiences that led him to Cognita, Chen shared:
I was a postdoctoral researcher, working on foundation models in radiology with my co-founders, Louis Blankemeier and Akshay Chaudhari, and my advisor Curt Langlotz. My research and engineering interests lie in representation learning in the multimodal domain, especially vision and language.
This interest took root during my junior year, when I began studying image captioning, the process of transforming unstructured images into meaningful text, for example, “A cat is sitting on top of a suitcase.”
Afterward, during my Ph.D., I turned to the mirrored problem in the medical domain: learning clinically meaningful representations from medical images and turning them into diagnostic sentences.
To me, it is almost a gaming experience to teach the machine to use continuous vectors to represent images and text. More fascinatingly, this “game” could potentially create real-world good.
Recently, working toward making a scalable impact, we started Cognita, where my role as CTO is, first, to create systems that can interpret medical images comprehensively to assist radiologists, and second, to create technical infrastructure that enables us to create such systems.
How Cognita Started
When discussing how the idea for Cognita came together, Chen explained:
The idea came from our desire to improve healthcare. Radiology naturally became our focus because it plays such a central role in clinical workflows, and it is also the area where we have spent years applying AI in our research.
For years, people have commonly believed that radiology AI is a solved problem. However, the reality is that most AI products are doing something much narrower than what radiologists do. Most companies train models that look for a single disease with a yes-or-no answer, while radiologists review the current study exam, imaging trajectory, and clinical history and describe the findings in detail, such as visual characteristics, location, size, severity, and uncertainty.
Essentially, this is not only a product issue but a solution issue. The problem is being solved in an unscalable way. We recognized that there was a huge opportunity to instead tackle the problem of drafting comprehensive results for radiologists.
The company really took shape through our early collaboration with Radiology Partners, the largest radiology practice in the U.S. As we worked closely with practicing radiologists to co-develop our models, it became increasingly clear that this technology could address very real problems, including growing backlogs, increasing imaging volumes, and limited access to imaging expertise in many regions.
Early Model Validation
When asked about his favorite memory from working at Cognita, Chen recalled:
One of my favorite moments is the first human evaluation after days and nights of tackling the technical challenges with our Cognita team in the beginning.
It was in a Zoom meeting with Louis and Sriyesh Krishnan, a clinical AI team member at Radiology Partners. In an evaluation batch, the model outputs were preferred over real reports in a blinded setting, validating our initial direction of the solution.
Core Products And Features
When asked about Cognita’s core products and features, Chen detailed:
At Cognita, we build foundation models that interpret medical images end-to-end and draft comprehensive radiology results. Today, those models focus on X-rays and CT scans, with the goal of supporting radiologists throughout the reporting process.
Most imaging AI on the market is designed to detect a single finding. That can be useful in narrow cases, but it doesn’t reflect how radiologists actually work. They’re responsible for synthesizing all relevant findings into a coherent report.
Our goal is to output comprehensive results that the radiologist reviews, edits, and signs. The model isn’t making final decisions. It’s helping reduce the time and effort required to get from image to report while keeping the physician in control.
Through our work with Mosaic Clinical Technologies at Radiology Partners, these models now power Mosaic Drafting, which is being evaluated on X-rays and CTs in clinical settings. The early feedback has been encouraging, particularly around usability and fit within existing workflows.
Handling Real-World Variability
When asked about challenges Cognita has faced in its sector, Chen noted:
Most of the challenges we’ve faced have been technical, largely stemming from translating research into models that are robust to real-world variability.
Radiology is defined by corner cases. To provide an extreme example, which might not be too far off, imagine that there are 10,000 rare diseases and each of these diseases occurs in one in 10,000 cases.
This illustrates how radiologists encounter rare diseases very often. To handle this variability, they need to pull from their decade of medical training and continue to learn throughout their careers.
It is very challenging to train models that can handle this.
FDA Breakthrough Device Designation
When asked about Cognita’s most significant milestones, Chen said:
One of the company’s most significant milestones was receiving FDA Breakthrough Device Designation for our chest X-ray generative AI model.
This was particularly meaningful because it represents the first time a generative radiology model has received this designation.
The designation allows us to work closely with the FDA as we move toward clearance and reinforces the potential of our models to address the growing radiology capacity bottleneck.
Team Differentiation
When asked what differentiates Cognita from its competition, Chen emphasized:
Probably one of our biggest differentiators is how we operate as a team.
Our team is small but agile, and we decided to keep it that way on purpose. With the critical need for solutions to the radiology problem, we wanted to have an environment where roles are not narrowly defined, and everyone at Cognita jumps in whenever there is a problem to solve.
Future Goals
When discussing Cognita’s future goals, Chen concluded:
Broadly, our mission consists of three stages.
First, we aim to train AI models that mimic what radiologists do today, and we are already well advanced in this phase.
Second, we want to add capabilities that go beyond current radiology practice, such as disease risk prediction and opportunistic imaging, to unlock the incredibly dense information contained in medical images that remains largely underutilized.
Third, we plan to extend our models beyond radiology to understand the vast swaths of healthcare signal data within the health system, including pathology, omics, lab values, and EHR data.
Together, these three stages move us toward our mission of significantly increasing the world’s access to healthcare. We believe this is the right path to achieve our mission of significantly increasing the world’s access to healthcare.
