A multi-agent AI system called PRISM reduced outright rejection of neuroendocrine tumor treatment recommendations from 18% to 8% by retrieving clinical evidence, assessing trial eligibility, and generating verified treatment recommendations, while large language models demonstrated 95% accuracy in selecting PET tracers and deep learning achieved 90% accuracy in predicting lung cancer biomarkers.
- PRISM multi-agent AI system reduced treatment recommendation rejection from 18% to 8% and improved composite quality scores from 0.77 to 0.84
- Treatment completeness increased from approximately 70% to nearly 90% with improved citation accuracy (4.33 vs 1.86 clinical trials per recommendation)
- LLMs achieved 95% accuracy in selecting PET tracers (FDG, PSMA, DOTATATE) from free-text clinical referrals across four different models
- Deep learning predicted PD-L1 expression in lung cancer with 0.90 AUC internally and 0.84 AUC externally, approaching fully supervised baseline performance with only 20% labeled data
- AI designed to support, not replace clinical decision-making, with locally deployed models offering better data protection than cloud-based alternatives
Vienna -- A multi-agent AI system reduced outright rejection of neuroendocrine tumor treatment recommendations from 18% to 8% in a blinded expert comparison, researchers reported at the European Association of Nuclear Medicine (EANM) 2026 congress in Vienna.
The study was among several presented during the Cutting Edge Science Track session "LLMs and Other Innovative DL Approaches" on Oct. 19, chaired by Laszlo Papp, PhD, of the Medical University of Vienna in Austria, and Lara Cavinato. Other presentations examined large language models (LLMs) for PET/CT protocol selection and deep learning for predicting lung cancer biomarkers.
Pia Koller, MSc, of Bern University Hospital and the University of Bern in Bern, Switzerland, presented PRISM, a seven-agent AI system designed to support multidisciplinary neuroendocrine tumor (NET) boards.
"Static guidelines take time to update, so they often lag behind practice-changing radiopharmaceutical trials," Koller explained.
PRISM addresses this gap by retrieving clinical evidence, assessing trial eligibility, generating treatment recommendations, and checking citations and potential errors before producing its final output.
The researchers tested the system on 15 complex cases from five European NET centers. Five international experts assessed its recommendations against those generated by a conventional single-prompt LLM.
PRISM improved the composite quality score from 0.77 to 0.84 (p < 0.01). Treatment completeness increased from approximately 70% to nearly 90%, while outright rejection fell from 18% to 8%.
The system also cited more than twice as many traceable clinical trials per recommendation, averaging 4.33 compared with 1.86 for the standard LLM. No fabricated references were reported, although numerical errors remained.
Interestingly, the additional transparency did not necessarily make experts more trusting. "Traceable citations let experts check the evidence and spot numerical errors," Koller said.
The study was limited to 15 cases evaluated offline. Koller emphasized that PRISM is intended to support, rather than replace, clinical decision-making.
Tracer selection
Dr. Clemens Backhaus of the Department of Nuclear Medicine at University Hospital Münster in Münster, Germany, presented a study assessing whether LLMs could select PET tracers and imaging protocols from free-text referrals.
The researchers tested GPT-o3, GPT-5, GPT-OSS 120B, and Llama 4 on 201 clinical requests, comparing their recommendations with those of three residents and a reference standard established by two nuclear medicine experts.
All four models achieved approximately 95% accuracy when choosing among FDG, PSMA, and DOTATATE tracers. Results were less consistent for the field of view.
GPT-o3 and GPT-5 performed comparably to residents, while the locally deployed models frequently recommended longer scan coverage than necessary. "Large language models are potentially a supportive tool for protocoling PET/CT examinations," Backhaus concluded.
The researchers also examined the trade-off between model performance and patient data protection. Locally deployed models can process identifiable patient information within hospital systems, whereas cloud-based models require additional safeguards.
The study included only three tracers and two field-of-view options. The models were tested without fine-tuning, and residents had the advantage of knowing their institution's protocols.
Predicting PD-L1 expression
Prof. Dimitris Visvikis, research director at the French National Institute of Health and Medical Research (Inserm) and the University of Brest in Brest, France, presented research by Ronrick Da-Ano and colleagues on predicting programmed death-ligand 1 (PD-L1) expression in non-small cell lung cancer.
The study, published in Scientific Reports in June 2026, combined semi-supervised and unsupervised learning to make use of PET/CT scans without PD-L1 labels.
The researchers used 384 patients for model development and 464 for external validation. The approach achieved an area under the curve (AUC) of approximately 0.90 internally and 0.84 externally.
With only 20% of labeled training data, the model approached the performance of a fully supervised baseline trained with 80%.
The researchers assessed PD-L1 expression at a threshold of 1%, rather than predicting immunotherapy response. The method remains experimental. "We do not have 100% prediction, so we do not say that it can replace immunohistochemistry," Visvikis said.
Our full coverage of EANM 2026 can be found here.
