Two experts explained how generative AI can support and optimize work processes in clinical radiology in a session at the European Congress of Radiology (ECR) in Vienna. The issue of sustainability was also the subject of lively discussion.
Marc Kohli from San Francisco, USA, gave an example of how large language models (LLMs) can help to simplify patient instructions. ‘In radiological practice, there are often thousands of different examinations with specific instructions – these are usually written by technologists or radiographers who are not always linguistically trained,’ said Kohli. He presented a case in which instructions for a CT colonoscopy were analyzed with LLMs and simplified so that the word count was reduced, foreign words were replaced, and the text became easier to understand.
Ambient Dictation
As Kohli explained, LLMs can also help optimize radiological reports. With ambient dictation, radiologists dictate their observations as they would to a colleague or trainee. The LLM automatically creates a structured report from this. However, in the discussion that followed the presentation, it was pointed out that training data for speech recognition is often male-dominated, which could affect the accuracy of female voices.
Literature research with LLMs
To illustrate how LLMs can also be used for literature research or topic identification, Kohli fed Google’s ‘Zero-Code RAG’ with his publications and had the main topics of his scientific work summarized. On request, the model also generated an overview of key ethical issues in AI, including specific citations and sources.
Kohli concluded, that the increasing complexity of diagnostics and the growing volume of image data make the use of intelligent decision-support systems indispensable.
Don’t neglect sustainability
In her presentation, Florence Xini Doo from Baltimore (USA) focused on the aspect of sustainability when using (generative) AI in radiology. She presented three approaches to measuring environmental impact: direct measurements using electricity meters, energy measurement software, and retrospective calculations based on the software’s runtime. Doo particularly emphasized the enormous energy and water consumption when training large AI models such as Chat GPT-3. ‘The cost of training such a model is estimated at 60 to 70 million dollars and an energy consumption of up to 1,287 megawatts – comparable to the emissions of 118 petrol-powered vehicles per year,’ said the expert. However, training in the cloud is significantly cheaper than in local data centers, which also require considerable amounts of drinking water for cooling. ‘The growing number of data centers is increasingly competing with agricultural and municipal water sources. By 2027, water consumption by AI will be four to six times higher than Denmark’s annual water consumption,’ Doo sounded the alarm.
Sustainability benchmarks for AI
She called for the development of sustainability benchmarks for AI in healthcare and the establishment of a ‘Radiology AI Eco-Label’ certification modeled on energy efficiency labels or LEED certifications. Doo also appealed to the audience to find innovative solutions for the sustainable use of AI in radiology and to work together towards a greener future.
At the end of the session, the participants agreed that radiologists are more open to AI solutions as they are already familiar with generative models from their everyday lives. This was also made clear by a survey of the audience. The rapid success of tools such as ChatGPT also suggests that acceptance of AI-supported reporting may be growing faster than previous technological changes. This was also underpinned by figures from the USA, where the use of AI rose from 33% in 2020 to 50% in 2023.
