Generated AI with large-scale language models such as ChatGpt is growing in industries such as customer service and creative content production. However, healthcare has moved more carefully.
Radiology, a specialist focused on digital image analysis and pattern recognition, has emerged as the forefront of adopting new AI technologies.
That's not to say that AI is new to radiology. The Department of Radiology was one of the most infamous AI predictions when Nobel Prize winner Jeffrey Hinton said in 2016, “People should stop training radiologists now.”
However, almost ten years later, AI conversion in the field is on a significantly different path. Although radiologists are not exchanged, they integrate the generated AI into their workflows to tackle labor-intensive tasks that do not require clinical expertise.
“As opposed to worrying about AI, radiologists want it to serve the challenges of the workforce,” explained Dr. Curt Langlotz, Senior Deputy Director of Research at Stanford.
Regulatory challenges for generation AI in radiology
Hinton's concept wasn't entirely out of style. Many radiologists now have access to predictive AI models that classify images or highlight potential anomalies. Langlotz said the rise of these tools “created an industry” for over 100 companies focusing on AI in medical imaging.
The FDA lists over 1,000 AI/ML-enabled medical devices. This can include algorithms and software. However, approved devices are based on more traditional machine learning techniques, rather than generator AI.
Ankur Sharma, Bayer's Medical Devices and Radiology Head of Medical Devices and Radiology, explained that the AI tools used in the radiology are categorized as computer-aided detection software that helps analyze and interpret medical images. Examples include triage, detection, and characterization. Each tool must meet regulatory criteria, including research to determine detection accuracy and false positive rates, among other metrics. This is particularly challenging for newer, less-understandable generation AI technologies.
Characterization tools that analyze specific anomalies and suggest what they are face the highest regulatory standards as both false positives and denials take risks. As Hinton envisioned, the kind of GEN AI radiologist idea that allows for automatic diagnosis is classified as “characterization” and must meet high standards of evidence.
Regulation is not the only hurdle that must be leapt to see wider use in radiology.
Today's best general purpose large-scale language models, like Openai's GPT4.1, are trained with trillions of data. Scaling the model this way gave me great results as the new LLM always beat the older model.
However, training radiology-generating AI models at this scale is difficult. This is because the amount of training data available is much smaller. Healthcare organizations also have no access to the access to calculate enough resources to build models at the scale of the largest large language model, and it costs hundreds of millions of dollars to train.
“The size of training data used to train the largest text or language models within medicine shows a 100-fold difference from external medicine,” says Langlotz. The biggest LLMS training rubs almost the entire internet. The medical model is limited to images and data that the agency can access.
Current reality of generation AI in radiology
These regulatory obstacles appear to question the usefulness of generative AI in radiology, particularly when making diagnostic decisions. However, radiologists find technology useful in their workflows as they can take on some of the daily labor-intensive management tasks.
For example, Sharma said some tools can take notes as radiologists direct them to observe medical images. Several large language models “take these reports and translate them into more patient-friendly languages,” he added.
Dr. Langlotz said the products that draft the report can provide radiologists with “substantial productivity benefits.” He compared it to having resident trainees drafting reports for review. This is a resource often available in academic settings, but not in radiology practices such as hospital radiology.
Sharma said that by automating and streamlining reporting, follow-up management, and patient communication, the generation AI can help radiologists and give radiologists time to focus more on “reading expertise” including image interpretation and diagnosis of complex cases.
For example, in June 2024, Bayer and Rad AI announced a collaboration to integrate generated 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 enable radiologists to create reports more efficiently using Rad AI technology. For example, RADAI can use the generated AI transcription to generate written reports based on the radiologist's instructions. Such applications do not directly affect diagnostics, thus reducing regulatory hurdles.
Looking ahead, Dr. Langlotz said he expects to see even greater adoption of AI in the near future. “I think there will be changes in the daily work of a radiologist over the next five years,” he predicted.
