Be the first to react To strengthen the attributes of large language models applied to radiology, the RAG, which code information in a vector space to refine the tasks of LLM based on knowledge, seems relevant. In a study published in the journal Radiology: Artificial Intelligence, researchers are testing a RAG on the best known of the LLM of radiological images.
Large language models (LLM) using an architecture based on machine learning can learn, synthesize and extract information in natural language in several areas, including diagnostic radiology. These models display impressive performance for various tasks, such as answering questions, modifying radiology reports and even generating accounting prints.
RAG to strengthen the attributes of large language models applied to radiology
But these models are likely to create hallucinations, generate constraints in terms of IT and economic resources and are on the other hand insecure regarding the confidentiality of patient data, which currently limits their general integration in clinical workflows.
The increased generation of recovery (Retrieval -Augmented Generation – RAG) is a strategy aimed at improving the performance of these large language models by providing them with an updated knowledge corpus that can be used for the generation of real -time responses. The relevant text of the knowledge base is combined with the user’s prompts in order to provide the best response based on this corpus of information, which could allow the RAG to respond to LLM concerns without requiring expensive settings, while serving as a clinical workflow.
Information coded in a vector space to refine the tasks of LLM based on knowledge
Technically, the RAG implies the orchestration of an information corpus in a vectorized format using a recovery method and a basic LLM. The information is coded in a large vector space using an integration model, which allows an effective search for similarity and the recovery of documents from a vector database storing integration indexes. The documents recovered are then provided at the LLM to serve as a context for the generation of the final model which allows, in the end, to improve knowledge based on knowledge, in particular the answer to simple questions, the generation of texts or open images.
Researchers test a RAG on the best known of the LLM of radiological images analysis
A retrospective study published in the journal Radiology: Artificial Intelligence looks at this problem. A radiology specific RAG system was developed using a vector database of 3,689 articles published in the journal radiographics from January 1999 to December 2023. The performance of five LLM with RAG-SYSTEMS and without RAG on a radiology exam of 192 questions were compared. Researchers have shown that RAG has significantly improved exam scores for GPT-4 and R+Command, but not for Claude Opus, Mixtral or Gemini 1.5 Pro.
RAG-SYSTEMS obtained significantly better results than LLM PURS on a subset of 24 questions coming directly from radiographics (85 % against 76 %). Rag-Systems has extracted 21 from the 24 relevant radiographic references cited in the explanations of the responses of the exam and successfully cited them in 18 of the 21 results. The results suggest that the RAG is a promising approach to improve the capacities of LLM for knowledge tasks in radiology, by offering a search for transparent information and specific to the field.
Royan Paolo
