Skip to main content
NCBI home page
RSNA Journals logoLink to RSNA Journals
letter
. 2025 Apr 1;315(1):e242454. doi: 10.1148/radiol.242454

Beyond Proprietary Models: The Potential of Open-Source Large Language Models in Radiology

Satvik Tripathi 1,✉, Ali S Tejani 2, Tessa S Cook 1
PMCID: PMC13316792  PMID: 40167433

Editor:

The recent article by Dr Adams and colleagues (1), in the August 2024 issue of Radiology, highlights a significant milestone in the evolution of large language models (LLMs), demonstrating that open-source models like Llama 3 are not only maturing rapidly but are also beginning to rival proprietary models such as GPT-4. The study revealed that Llama 3 achieved a comparable accuracy rate to GPT-4 on American College of Radiology (ACR) in-training test questions, with an overall accuracy of 74% compared with 78% for GPT-4, and on radiology board–style questions, with an accuracy of 80% compared with 82% for GPT-4. These results challenge the traditional preference for proprietary models, which have historically been favored due to their extensive training on large, curated datasets and their robust computational backing (2,3).

The results of the study should be interpreted cautiously, as high performance on ACR in-training test questions might be due to data leakage from publicly available questions. Moreover, it raises the question of whether such examinations are a meaningful benchmark for model performance. Does good performance reflect the model’s ability, or does it expose flaws in the questions? However, the authors set a strong example for future LLM testing by using new radiology board–style questions to ensure assessment on unseen content and by evaluating performance across multiple runs to address concerns about model stochasticity.

With competitive performance, open-source models offer several advantages that are particularly relevant for the future of artificial intelligence (AI) in radiology. First, they are more accessible to the global research community, enabling widespread experimentation and adaptation to specific clinical needs. Second, their transparency allows for greater scrutiny and refinement, which can lead to more reliable and reproducible AI tools. Third, unlike proprietary models, where modifications and retraining are often limited or costly, open-source models allow researchers to continuously iterate and optimize the model for specific medical applications.

The radiology community must now consider the broader implications of these findings. Open-source models hold the potential to democratize AI in health care, fostering greater collaboration and knowledge sharing across institutions. Furthermore, the transparency and accessibility of open-source AI can enhance the reproducibility of research, a critical factor in validating the robustness of AI tools in clinical settings.

Footnotes

Disclosures of conflicts of interest: S.T. No relevant relationships. A.S.T. No relevant relationships. T.S.C. Grants to institution from the National Institutes of Health, Independence Blue Cross, RSNA, and American College of Radiology; consulting fee from Cuattro; honoraria for lectures from ISMIE, Icahn School of Medicine, and Massachusetts General Hospital; support for travel expenses as a board member to annual retreat or relevant meetings from the Society for Imaging Informatics in Medicine and Pennsylvania Radiological Society; chair of the Society for Imaging Informatics in Medicine (2022–2024); president of the Radiology Alliance for Health Services Research (2023–2024); alternate councilor for Pennsylvania Radiological Society; member of the Pennsylvania Radiological Society Executive Committee (ended 2023); and board member for the Association of University Radiologists (ended 2024).

References

  • 1. Adams LC , Truhn D , Busch F , et al . Llama 3 challenges proprietary state-of-the-art large language models in radiology board–style examination questions . Radiology 2024. ; 312 ( 2 ): e241191 . [DOI] [PubMed] [Google Scholar]
  • 2. Bhayana R , Krishna S , Bleakney RR . Performance of ChatGPT on a radiology board–style examination: insights into current strengths and limitations . Radiology 2023. ; 307 ( 5 ): e230582 . [DOI] [PubMed] [Google Scholar]
  • 3. Krishna S , Bhambra N , Bleakney R , Bhayana R . Evaluation of reliability, repeatability, robustness, and confidence of GPT-3.5 and GPT-4 on a radiology board–style examination . Radiology 2024. ; 311 ( 2 ): e232715 . [DOI] [PubMed] [Google Scholar]
Radiology. 2025 Apr 1;315(1):e242607. doi: 10.1148/radiol.242607

Response

Lisa C Adams 1,✉, Keno K Bressem 1,2

We appreciate the thoughtful letter by Dr Tripathi and colleagues regarding our recently published article on the performance of open-source LLMs on radiology board–style examination questions. Their concern about data leakage is valid; while such leakage could potentially inflate performance metrics, it likely affected all models equally, preserving the relative performance differences. Still, to mitigate this issue, we included 85 new radiology board–style questions and evaluated performance across multiple runs, addressing concerns about model stochasticity. These approaches underscore the need for new datasets and benchmarks for future evaluations and set a precedent for robust LLM testing in medical domains.

Beyond the issue of data leakage, multiple-choice questions, while useful for initial benchmarking, might not be the optimal benchmark for evaluating clinical utility and can provide unreliable estimates of model performance (1). Multiple-choice questions offer a limited view of models’ capabilities, particularly in the context of complex clinical decision-making. Our study was a preliminary exploration, and we acknowledge the need for more rigorous, clinically focused evaluations to better assess the potential of these models in health care settings.

Recent efforts have aimed to develop benchmarks that emphasize clinical utility over mere knowledge recall. For example, the LongHealth benchmark assesses LLMs using detailed clinical documents, challenging them with tasks that reflect real-world scenarios (2). Similarly, the Clinical Language Understanding Evaluation, or CLUE, benchmark offers a comprehensive evaluation of LLMs in practical clinical applications, moving beyond traditional multiple-choice questions (3).

Additionally, we support the perspective that LLMs should be viewed as reasoning engines rather than mere repositories of knowledge (4). This aligns with emerging applications in radiology, such as using LLMs for post hoc transformation of free-text radiology reports into structured formats. Our recent work demonstrated the capability of GPT-4 in this area, highlighting its potential to enhance data management and standardize reporting practices across different languages and institutions (5).

In conclusion, while our study provides an initial insight into the capabilities of open-source LLMs in radiology, further development of clinically relevant benchmarks and evaluation methods is crucial. Recognizing both the potential and the limitations of these models will be essential as we work toward their responsible integration into clinical practice.

Footnotes

Disclosures of conflicts of interest: L.C.A. No relevant relationships. K.K.B. Grants from the European Union (101079894), Bayern Innovativ, German Federal Ministry of Education and Research, and Wilhelm Sander Foundation; payment or honoraria for lectures, presentations, speakers bureaus, manuscript writing, or educational events from Canon Medical Systems and GE HealthCare; and advisor for the European Union Horizon 2020 LifeChamps Project (875329) and the European Union Innovative Health Initiative IMAGIO Project (101112053).

References

  • 1. Pezeshkpour P , Hruschka E . Large language models sensitivity to the order of options in multiple-choice questions . arXiv 2308.11483 [preprint] https://arxiv.org/abs/2308.11483. Posted August 22, 2023. Accessed January 4, 2025 .
  • 2. Adams L , Busch F , Han T , et al . LongHealth: A Question Answering Benchmark with Long Clinical Documents . arXiv 2401.14490 [preprint] https://arxiv.org/abs/2401.14490. Posted January 25, 2024. Accessed January 4, 2025 . [DOI] [PMC free article] [PubMed]
  • 3. Dada A , Bauer M , Contreras AB , et al . CLUE: a clinical language understanding evaluation for LLMs . arXiv 2404.04067v3 [preprint] https://arxiv.org/abs/2404.04067v3. Posted April 5, 2024. Updated June 24, 2024. Accessed January 4, 2025 .
  • 4. Truhn D , Reis-Filho JS , Kather JN . Large language models should be used as scientific reasoning engines, not knowledge databases . Nat Med 2023. ; 29 ( 12 ): 2983 – 2984 . [DOI] [PubMed] [Google Scholar]
  • 5. Adams LC , Truhn D , Busch F , et al . Leveraging GPT-4 for post hoc transformation of free-text radiology reports into structured reporting: a multilingual feasibility study . Radiology 2023. ; 307 ( 4 ): e230725 . [DOI] [PubMed] [Google Scholar]

Articles from Radiology are provided here courtesy of Radiological Society of North America

RESOURCES