University of California, San Francisco (UCSF) radiology residents Yannan Yu, MD, and Michael Romano, MD, PhD, presented during the recent UCSF Center for Intelligent Imaging (ci2) SRG Pillar Meeting. Yu presented her research on using large language models (LLMs) to generate multilingual, patient-friendly radiology reports, while Romano introduced the first publicly available multimodal imaging and genomics resource for primary central nervous system lymphoma (PCNSL).
Radiology reports are written for clinicians, with an average reading level above 12th grade, while the average American adult reads between a seventh- and eighth-grade level. More than 25 million people in the United States also have limited English proficiency, and while in-person and video interpreters are available during clinical visits, no equivalent exists for written radiology reports.
Yu and her team designed a three-phase study. They first tested prompting strategies and found that a structured prompt with a radiologist-supervised retrieval pipeline produced the most accurate and readable output. A radiologist reader study then evaluated 250 simplified reports in English, Chinese and Spanish, showing high accuracy, strong readability and very low error rates. A patient survey of 29 participants followed, revealing significant improvements in understanding and confidence, along with reduced anxiety. Of those surveyed, 96.4% said including simplified reports alongside the originals would be beneficial.
Romano tackled a different gap in radiology. PCNSL is a rare non-Hodgkin lymphoma with an incidence of 0.7 per 100,000 in the United States and a median survival of just over two years. Its variable appearance on MRI makes prospective diagnosis difficult, and biopsy remains necessary for definitive diagnosis. No public dataset previously existed for researchers building computational models around this disease.
Using the UCSF de-identified clinical data warehouse, Romano and his collaborators curated data from 150 patients. The dataset includes skull-stripped MRIs across four sequences, lesion segmentation masks, radiomics features, clinical data, and UCSF500 genomic panel results for 64 of those patients. The team used PyAlpha for skull stripping and lesion segmentation, then manually adjusted the masks before extracting radiomic features. They also built an API and analysis tools so other researchers can work with the data. Among the initial findings, the MYD88 mutation appeared in 73% of the genomic cohort, consistent with known PCNSL genomic profiles. Romano noted that the dataset could support future work in early detection, genomic prediction, and treatment response analysis.
Visit the ci2 events page to learn more about upcoming SRG Pillar meetings.
