Purpose or Learning Objective: Optimizing magnetic resonance imaging protocoling for sports injuries is crucial
for image quality and accurate diagnoses. Manually navigating guidelines for protocol
selection is time consuming and prone to error. Large language models offer potential
automation but may lack training on region-specific magnetic resonance imaging guidelines.
Retrieval-augmented generation–enabled models could address this limitation by integrating
external knowledge into large language model prompts. This study evaluates the performance
of a baseline and retrieval-augmented generation–enabled large language model in generating
magnetic resonance imaging protocols for musculoskeletal sports injuries.
Methods or Background: The European Society of Musculoskeletal Radiology 2016 Guidelines for MR Imaging
of Sports Injuries were used to develop a custom architecture that retrieves protocol
parameters for 19 body regions of the musculoskeletal system. GPT-4o was prompted
with and without retrieval-augmented generation to produce guideline-based recommendations
for each magnetic resonance imaging sequence. The evaluated parameters included field
of view (max), slice thickness (max), echo time, and matrix size (min). Additionally,
both models provided patient positioning recommendations. Completeness and adherence
to European Society of Musculoskeletal Radiology guidelines were assessed, with adherence
rates compared between the retrieval-augmented generation and non–retrieval-augmented
generation models. Subgroup analyses were conducted by magnetic resonance imaging
parameter and body region.
Results or Findings: A total of 109 magnetic resonance imaging sequences were identified from the European
Society of Musculoskeletal Radiology guidelines, yielding 436 parameters for evaluation.
Both models achieved 100% completeness in parameter output generation. Adherence rates
were 97.5% for the retrieval-augmented generation model versus 31.9% for the non–retrieval-augmented
generation model. McNemar's test (χ2 = 278.2; P < 0.001) indicated a significant difference between models. Subgroup analyses showed
significant improvements across all parameters (P < 0.05), with the largest increases in slice thickness (98.2% versus 25.5%) and matrix
size (100% versus 32.1%).
Adherence also significantly improved across all body regions, with the greatest improvement
observed for spine sequences (80% versus 0%), and the smallest for the shoulder (77.5%
versus 54.2%). For patient positioning recommendations, the retrieval-augmented generation
model achieved 94.7% adherence versus 68.4% for the non–retrieval-augmented generation
model (P < 0.05).
Conclusion: This is the first study to develop a custom retrieval-augmented generation–enabled
system for magnetic resonance imaging protocoling in musculoskeletal radiology. Retrieval-augmented
generation–based prompts significantly improved adherence to magnetic resonance imaging
guidelines across parameters, body regions, and positioning recommendations, demonstrating
their potential for workflow optimization. Future work should focus on prospective
clinical validation and integration with radiology communication systems.
