Australia-based startup Harrison.ai spent the past four years building out AI-powered medical diagnostic software and services. The health tech company is focused on using AI to improve the speed and accuracy of radiology image analysis and enhance diagnostic accuracy in pathology.
This week, the company has unveiled what it considers to be a "groundbreaking" radiology-specific vision language model designed to excel at radiology tasks. It's part of the startup's broader goal to use AI automation to scale the global capacity of healthcare.
Harrison.rad.1 is a foundational model that is dialogue-based. It can answer open-ended visual questions about radiology images, detect findings, provide descriptions, and generate structured reports, providing longitudinal reasoning based on clinical history and patient context, according to executives.
"It is the most performant radiology-specific foundational model that is going to be available to date," Aengus Tran, M.D., co-founder and CEO of Harrison.ai said in an exclusive interview with Fierce Healthcare. "What makes the model truly different is that unlike other generic, large language models out there that have been trained to perform general tasks, Harrison.rad.1 is a radiology-specific foundation model tuned to excel at radiology tasks and prioritize factual correctness and clinical helpfulness."
The model is now being made accessible to selected industry partners, healthcare professionals and regulators to open up the conversation on the responsible use of AI in healthcare, Tran noted.
Many existing generative AI models are functionally generic and predominantly trained on general and open-source data. Harrison.rad.1 has been trained on real-world, diverse and proprietary clinical data, comprising millions of images, radiology studies and reports. The dataset is further annotated at scale by a large team of medical specialists to provide Harrison.rad.1 with clinically accurate training signals, according to executives.
Harrison.ai's LLM has outperformed other foundational models in benchmarks, according to the company. Harrison.rad.1 demonstrated notable performance, excelling in radiology examinations designed for human radiologists. Specifically, it surpasses other foundational models on the challenging Fellowship of the Royal College of Radiologists (FRCR) 2B Rapids examination—an exam that only 40% to 59% of human radiologists manage to pass on their first attempt. When reattempted within a year of passing, radiologists score an average of 50.88 out of 60. Harrison.ai performed on par with accredited and experienced radiologists at 51.4 out of 60, while other competing models such as OpenAI’s GPT-4o, Anthropic’s Claude-3.5-sonnet, Google’s Gemini-1.5 Pro and Microsoft’s LLaVA-Med scored below 30 on average.
“AI’s promise rests on its foundations—the quality of the data, rigor of its modeling and its ethical development and use. Based on these parameters, the Harrison.rad.1 model is groundbreaking," Tran said.
The foundational language model builds on Harrison.ai's work over the past four years to develop and commercialize AI tools that support clinical diagnosis.
"We are on a mission to scale the global capacity of diagnosis and treatment and improve patient outcomes by building a suite of AI tools that free doctors from doing repetitive and mundane tasks so they can truly focus on what is uniquely human," Tran said.
In a short demo of Harrison.rad.1 using a chest X-ray, Tran showed how the technology can assist in radiology tasks when interpreting medical images. "If you look at this chest X-ray, one of the questions that people may ask is, 'Is this an urgent study requiring priority attention?' So you can ask Harrison.rad.1 and the model, through a dialog-based interface, would be able to provide a natural language answer to these clinical questions posed to it by the clinicians or radiologists," he said.
In the demo, the model flagged that the case was urgent based on the X-ray image. "There is a mass in the left lung and it has provided a helpful report to the user. This work would have otherwise taken several minutes from a trained radiologist in order to complete. The model itself is able to do this, of course, very accurately, but in a much shorter period of time. The idea is that this technology would empower the next generations of AI solutions in healthcare to effectively increase the available capacity of diagnostic tests that is available to clinicians in the United States but also across the world," he said.
Companies have developed machine learning algorithms to aid with image analysis, but Tran said Harrison.rad.1 represents a "quantum leap" to expand that capability.
"If you take an analogy of a spell checker, up to this point, the AI solution in medical imaging analysis has been very narrow. It's like a spell checker that's only checked for a few words, starting with maybe the letter 'A.' Up until this point, it's been quite a narrow solution. Think about [this solution] as a complete digital radiologist, rather than a point solution that checks for findings," he said.
Global healthcare is facing multiple intersecting challenges. There is a shortage of radiologists and pathologists along with rising imaging volumes and associated data per case. In the U.S. there are 11 radiologists per 100,000 people. More than two-thirds of the world’s pathologists are distributed across only 10 countries.
These shortages can lead to longer wait times for diagnostic results or more errors introduced by people working faster than ever, Tran said.
Health tech companies see the opportunity to use AI automation to help scale out clinical capacity. "The need for this technology has never been more acute," he noted.
Harrison.ai's existing radiology solution, Annalise.ai, has been cleared for clinical use in over 40 countries and is commercially deployed in healthcare organizations globally. The company is also developing a pathology solution, called Franklin.ai, that streamlines the microscopic examination of human tissues to identify signs of diseases, prognostic factors and markers that guide treatments.
"Harrison.rad.1 is the foundational model that will be powering all the solutions going forward to improve accuracy and functionality over time," Tran said.
The company's work in AI comes as the industry shifts from using AI as a research or innovation tool to introducing it more into clinical practice, he noted. "Fast forward five years, I truly believe that AI will play a pivotal role in solving capacity constraint in the United States healthcare system, and that ultimately will bring about more affordable and more equitable access to healthcare in the U.S."
How the startup approaches healthcare AI
Tran and his brother Dimitry Tran founded Harrison.ai in 2018 to build AI tools to support clinical diagnosis. Tran, a medical doctor by training, brought frontline clinical expertise to the startup while his brother Dimitry is a medical technologist and former Head of Innovation at Ramsay Health Care, one of the largest hospital groups in the world. Harrison.ai now has more than 200 employees globally.
The company's Annalise.ai solution is already commercially deployed across a number of healthcare systems globally, including across the NHS in the UK. The company is on track to process more than 7.5 million radiology scans in real-world clinical practice this year.
Annalise.ai detects up to 124 findings on chest X-rays and up to 130 findings on non-contrast head CTs and is customizable to an organization’s or healthcare systems’ needs, according to the company. Annalise’s chest X-ray (CXR) AI solution has been shown to improve diagnostic accuracy by 45% and efficiency by 12%, as validated by independent tests and recognized in international journals.
Harrison.ai has attracted the attention of leaders in the tech and healthcare worlds. Tesla’s Board Chair, Robyn M. Denholm, and Keith Dreyer, M.D., Mass General and Brigham‘s Chief Data Science Officer and Vice Chairman of Radiology, are both board members.
The startup has raised $141 million to date and is backed by Horizons Ventures, Blackbird Ventures, Skip Capital, Sonic Healthcare and I-MED Radiology Network. The company also has a partnership with Sonic Healthcare, one of the world’s largest medical diagnostics providers, to develop and commercialize new clinical AI solutions in pathology.
And, Harrison.ai's partnership with I-MED Radiology Network, the third largest private network of radiology clinics in the world, supported its work to build a radiology-specific foundation model. Through its partnerships with healthcare organizations, the company can leverage large data sets to build its clinically-focused technology.
Tran contends that developing its healthcare AI technology in Australia has given the company a leg-up in the AI space. In the U.S., FDA regulations make it more difficult for radiology AI companies to gain clearances for their technology. Outside the U.S., Harrison.ai has 254 findings approved with regulators such as TGA (Australia) or CE MDR (Europe) for its Annalise.ai solution. The company currently has 12 FDA-cleared findings.
"We are seeing a major innovation gap that is taking place, especially in the space of AI for diagnostic medicine, that is largely driven by the regulation by the FDA, who are focusing on regulating AI solution with validations and clearance criteria well above and beyond what the rest of the world is accepting. Because of that, most radiology AI companies need to take their very powerful and feature-full solution and really bring it down to only a numbers of functionality that the FDA would be willing to clear. That pulls over a 10x innovation gap between functionality and features that AI companies are able to deliver outside the U.S. versus inside the U.S.," he noted.
Harrison.rad.1 model will undergo further open and competitive evaluations by world-leading professionals, Tran said, underscoring the company's commitment to responsible AI development.
Harrison.ai wants to collaborate to speed up the development of further AI products, executives said. The company is making its foundational model available to researchers, industry partners and regulators to foster conversations about responsible AI, Tran said.
"The model itself is quite capable and broad in its capabilities and functionality and we believe that new methods of validation need to be created in order to best characterize the performance of the models and how it should be responsibly introduced into clinical practice," he said. "We're having multiple conversations with partners who have been reaching out to us on different products idea that what they want to build on top of the foundational model, but that is something that we will continue to explore once we understand the appetite in the industry."
Harrison.ai's plan is eventually to make the model commercially available for a number of use cases and to other organizations to build solutions on top of the company's technology. "The objective, right now, is to openly and competitively test the performance of this model for clinical accuracy and usefulness, and then we will be very keen to look at some of that application and use cases and start empowering those applications commercially in the near future," Tran said.