We utilize open-access LLMs to make radiology reports concise and well-structured and use Ollama platform to run LLMs locally behind a secure firewall.
We explore five different prompting approaches:
-
Structure: This approach focused solely on prompting the LLM to structure reports according to a predefined format.
-
Structure >> Conciseness: The LLM was initially prompted to structure the reports according to a specified format. Subsequently, another prompt instructed the model to refine the structured output further for conciseness. Here, each report processing consisted of two calls to the LLM.
-
Conciseness >> Structure: The LLM was first prompted to generate concise reports. Following this, the model was instructed to organize the concise information into a predefined structure. Here, each report processing consisted of two calls to the LLM.
-
Structure + Conciseness: The LLM was prompted to structure reports in a predefined format while also emphasizing the need for conciseness.
-
Structure + Conciseness (F, I): Given two prompts, the LLM was prompted to structure the Findings and Impressions sections in reports separately in a predefined format while also emphasizing the need for conciseness.