AI is set to continue making inroads in radiology in the coming years, according to two executives from Bayer and Google.

In an interview at Google Cloud Next, Bayer’s Guido Mathews and Google Cloud’s Shweta Maniar highlighted the transformative influence of the technology on the radiologist’s workflow, the increasing integration of AI into radiological education, and its potential to mitigate burnout and reduce error rates.

Bayer offers contrast agents and injectors for major radiology modalities, including CT, MRI, an angiography.

In addition to focusing on radiology, Bayer Pharmaceuticals is using generative AI models like Google Cloud’s Vertex AI and Med-PaLM 2 to streamline drug development. Bayer is also using Google’s high-performance computing resources for quantum chemistry calculations.

AI is fueling a radiological reboot

In 2016, deep learning pioneer Geoffrey Hinton predicted that AI systems would outperform radiologists by 2021, making them obsolete.

“We should stop training radiologists now,” he said.

The prediction has proven far from accurate. In many parts of the world, a shortage of radiologists remains a more vexing concern. In the U.S., 82% of practicing radiologists are 45 or older, according to the American College of Radiology. More than half are above the age of 55.

Meanwhile, demand for radiological imaging has expanded in recent years, fueling burnout and the potential for errors. Faced with often overwhelming workloads, the day-to-day radiologist error rate hovers in the range of 3% to 5%, according to a 2016 study in Insights into Imaging.

“In night shifts, especially, we see challenges,” said Mathews, the head of imaging, data and AI Research Center of Excellence at Bayer. “You miss things, right?”

A shifting perspective

Against this backdrop, AI’s role in radiology has shifted from an outright replacement to a valuable ally to overworked radiologists.

“Initially, there was a sentiment that we no longer needed radiologists,” said Mathews.

Now, the understanding is more that radiologists who don’t stay current with AI will fall behind, he argued.

Regulatory action has helped spur adoption. FDA has cleared hundreds of healthcare AI algorithms, most related to imaging.

The growing use of AI in radiology highlights the importance of healthcare companies like Bayer partnering with tech giants like Google. As the field continues to evolve, “it’s imperative for partners like Google and us to establish a framework for responsible AI, one that aligns with regulatory standards and meets bioethical requirements,” Mathews said.

The rise of AI-assisted diagnostic tools

Shweta Maniar, Google Cloud’s strategy and market leader responsible for biopharma in healthcare and life sciences, also highlighted the evolving role of AI in radiology, pointing to the growing use of AI-assisted diagnostic tools for triaging and other applications.

“In the education system, medical students are now using AI as part of their training,” she said. “So when they graduate, adopting AI won’t be an afterthought but a foundational aspect of their daily practice.”

Maniar underscored that radiologists are retaining control as AI makes headway in interpreting radiological imaging. “It’s a triaging opportunity so that the human in the loop is the one reviewing what needs attention.”

Patients also stand to benefit from the rise of AI in radiology, Mathews said.

“With industry partners like Google, we hope to make a significant difference, aiming to bring diagnostics to everyone globally and serve our patients better,” he said. “On the other hand, we need to collaborate within the industry, along with regulatory authorities. This collaboration is necessary to ensure the widespread availability of explainable AI and to introduce technology that, despite the human error rates we also see, but to have technology that is fulfilling its purpose and is very accurate.”

Retraining physicians — and AI

Continuing education is a fact of life for physicians.

“We also need to answer the question of how we will retrain AI, especially depending on its model, whether it’s a supervised learning system or whatever it is,” Mathews said. “How do we continually evaluate and update these technologies? I believe that while we need to develop technical frameworks, we must also establish regulatory frameworks to facilitate this process.”

The growing sophistication of AI frameworks will be instrumental to the process, Maniar said. Google’s Vertex AI, for instance, has emerging capabilities related to interpreting images, understanding speech-to-text translation and interpreting ambient documentation.

“With Vertex, we’ve seen customers play with AI for quite some time,” Maniar said. “Now we see customers starting to use Vertex AI and build their own solutions.”

AI continues to demonstrate the ability to streamline processes. Maniar said the creation of customized chatbots and semantic search applications may not be the flashiest of innovations, but their impact in terms of streamlining documentation-related tasks is undeniable.

Maniar also highlighted the launch of Med-PaLm 2, a large language model tailored for the medical domain. Google developed the model to accurately and safely answer medical questions. Earlier this year, Google Cloud announced that the model achieved “expert” test-taker level performance on the MedQA dataset of the U.S. Medical Licensing Examination (USMLE)-style questions with 85%+ accuracy. The company announced earlier this year that Med-PaLm 2 is the first AI system to score a passing mark on the MedMCQA dataset:

“We just announced at [Google Cloud Next] that Med-PaLm 2 is not only for trusted testers, but it will be open for a preview for life sciences and healthcare customers,” Maniar said. “These are the big categories of generative AI that organizations like Bayer can leverage.”