Back to all papers

An accessible efficient intervertebral disc vision-language model training strategy with LLM-based retrieval method for spine MRI report generation.

November 5, 2026pubmed logopapers

Authors

Chen X,Wang P,Wang Y,Han H,Ji X,Wang T,Wei B

Affiliations (7)

  • Center for Medical Artificial Intelligence, Shandong University of Traditional Chinese Medicine, Qingdao 266112, Shandong, China. Electronic address: [email protected].
  • Center for Medical Artificial Intelligence, Shandong University of Traditional Chinese Medicine, Qingdao 266112, Shandong, China. Electronic address: [email protected].
  • Department of Orthopaedic Surgery, The Affiliated Hospital of Qingdao University, Qingdao 266112, Shandong, China. Electronic address: [email protected].
  • Department of State-owned Assets and Laboratory Management, Shandong University of Traditional Chinese Medicine, Jinan 250355, Shandong, China. Electronic address: [email protected].
  • Department of Medical Imaging Center, Qilu Hospital of Shandong University Dezhou Hospital, Dezhou 253000, Shandong, China. Electronic address: [email protected].
  • Department of Orthopaedic Surgery, The Affiliated Hospital of Qingdao University, Qingdao 266112, Shandong, China. Electronic address: [email protected].
  • Center for Medical Artificial Intelligence, Shandong University of Traditional Chinese Medicine, Qingdao 266112, Shandong, China. Electronic address: [email protected].

Abstract

Artificial intelligence (AI) has become increasingly important in healthcare, and multimodal large language models (MLLMs) have demonstrated strong capabilities in understanding and integrating information from medical images and clinical text. This study explores a low-cost and effective MLLM-based approach for automated radiology report generation from lumbar spine MRI, aiming to improve reporting efficiency and consistency. To mitigate the challenges of limited data availability and the high costs associated with applying MLLMs to spinal report generation, we proposed using the axial plane of a single lumbar intervertebral disc (IVD) and the adjacent vertebrae as the research subject. We constructed the pipeline of data processing to get disc-corresponding images and medical reports based on large language models (LLMs). Then, We fine-tuned presently mature MLLMs under conditions of 1125 patient samples and hardware limitations, achieving considerable results. In addition to performance evaluated using traditional natural language processing metrics, including BLEU (0.24), ROUGE-L (0.53), and CIDEr (1.28), we further validated the proposed strategy on clinical pathological classification tasks, achieving F1 scores of 0.93 for disc pathology, 0.81 for canal stenosis, and 0.91 for dehydrated disc. Notably, the model achieved diagnostic performance comparable to that of experienced radiologists for canal stenosis and dehydrated disc. These results demonstrate the feasibility of the proposed strategy. Furthermore, our experiments show that the LLM-based data evaluation pipeline is stable and reliable. This study offers a new perspective on the field of spinal reporting and outperforms existing methods on both performance and cost. Meanwhile, there is a considerable room for improvement in aspects of data volume and training scale.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.