AI isn't replacing radiologists. Instead, they're using it to tackle time-sucking administrative tasks.

Generative AI powered by large language models, such as ChatGPT, is proliferating in industries like customer service and creative content production. But healthcare has moved more cautiously.

Radiology, a specialty centered on analyzing digital images and recognizing patterns, is emerging as a frontrunner for adopting new AI techniques.

That's not to say AI is new to radiology. Radiology was subject to one of the most infamous AI predictions when Nobel Prize winner Geoffrey Hinton said, in 2016, that "people should stop training radiologists now."

But nearly a decade later, the field's AI transformation is taking a markedly different path. Radiologists aren't being replaced, but are integrating generative AI into their workflows to tackle labor-intensive tasks that don't require clinical expertise.

"Rather than being worried about AI, radiologists are hoping AI can help with workforce challenges," explained Dr. Curt Langlotz, the senior associate vice provost for research and professor of radiology at Stanford.

Regulatory challenges to generative AI in radiology

Hinton's notion wasn't entirely off-base. Many radiologists now have access to predictive AI models that classify images or highlight potential abnormalities. Langlotz said the rise of these tools "created an industry" of more than 100 companies that focus on AI for medical imaging.

The FDA lists over 1,000 AI/ML-enabled medical devices, which can include algorithms and software, a majority of which were designed for radiology. However, the approved devices are based on more traditional machine learning techniques, not on generative AI.

Ankur Sharma, the head of medical affairs for medical devices and radiology at Bayer, explained that AI tools used for radiology are categorized within computer-aided detection software, which helps analyze and interpret medical images. Examples include triage, detection, and characterization. Each tool must meet regulatory standards, which include studies to determine detection accuracy and false positive rate, among other metrics. This is especially challenging for generative AI technologies, which are newer and less well understood.

Characterization tools, which analyze specific abnormalities and suggest what they might be, face the highest regulatory standards, as both false positives and negatives carry risks. The idea of a kind of gen AI radiologist capable of automated diagnosis, as Hinton envisioned, would be categorized as "characterization" and would have to meet a high standard of evidence.

Regulation isn't the only hurdle generative AI must leap to see broader use in radiology, either.

Today's best general-purpose large language models, like OpenAI's GPT4.1, are trained on trillions of tokens of data. Scaling the model in this way has led to superb results, as new LLMs consistently beat older models.

Training a generative AI model for radiology at this scale is difficult, however, because the volume of training data available is much smaller. Medical organizations also lack access to compute resources sufficient to build models at the scale of the largest large language models, which cost hundreds of millions to train.

"The size of the training data used to train the largest text or language model inside medicine, versus outside medicine, shows a one-hundred-times difference," said Langlotz. The largest LLMs train on databases that scrape nearly the entire internet; medical models are limited to whatever images and data an institution has access to.

Generative AI's current reality in radiology

These regulatory obstacles would seem to cast doubt on generative AI's usefulness in radiology, particularly in making diagnostic decisions. However, radiologists are finding the technology helpful in their workflows, as it can undertake some of their daily labor-intensive administrative tasks.

For instance, Sharma said, some tools can take notes as radiologists dictate their observations of medical images, which helps with writing reports. Some large language models, he added, are "taking those reports and translating them into more patient-friendly language."

Dr. Langlotz said a product that drafts reports can give radiologists a "substantial productivity advantage." He compared it to having resident trainees who draft reports for review, a resource that's often available in academic settings, but less so in radiology practices, such as a hospital's radiology department.

Sharma said that generative AI could help radiologists by automating and streamlining reporting, follow-up management, and patient communication, giving radiologists time to focus more on their "reading expertise," which includes image interpretation and diagnosis of complex cases.

For example, in June 2024, Bayer and Rad AI announced a collaboration to integrate generative AI reporting solutions into Bayer's Calantic Digital Solution Platform, a cloud-hosted platform for deploying AI tools in clinical settings. The collaboration aims to use Rad AI's technology to help radiologists create reports more efficiently. For example, RadAI can use generative AI transcription to generate written reports based on a radiologist's dictated findings. Applications like this face fewer regulatory hurdles because they don't directly influence diagnosis.

Looking ahead, Dr. Langlotz said he foresees even greater AI adoption in the near future. "I think there will be a change in radiologists' day-to-day work in five years," he predicted.