Leveraging Large Language Models for Multi-Modal Integration in Brain Tumor Diagnosis and Prognosis

A research team led by Dr. Zhuoqi Ma from Brown University has developed a method that integrates radiology and pathology reports using a pre-trained large language model (LLM) to improve the diagnosis and treatment of brain tumors, particularly gliomas. This innovative pipeline analyzed data from 426 patients, achieving a micro F1-score of 0.849 for tumor presence and 0.929 for tumor stability—significantly outperforming traditional single-source methods by over 10%. The LLM’s ability to synthesize diverse information enhances diagnostic accuracy and predicts survival outcomes without further training. The study identifies that predictions regarding tumor stability effectively differentiate high-risk from low-risk glioblastoma patients, rivaling established biomarkers like MGMT methylation status. Published in the KeAi journal Meta-Radiology, this research lays the groundwork for future oncology studies and suggests that incorporating modalities such as MRI and genomics could further advance personalized cancer care.

Source link