Can advanced large language models support radiology training? A performance assessment of DeepSeek R1
Thomas Saliba, Jacopo Ferrari, Chiara Pozzessere, David C. Rotzinger, Guillaume Fahrni
Abstract
<h2>Abstract</h2><h3>Background</h3> Large language models (LLMs) are increasingly used in medical education, including radiology training. DeepSeek-R1, an open-access LLM, has gained attention for its performance and accessibility. This study assesses DeepSeek-R1's ability to answer radiology-related questions based on the European Training Curriculum for Radiology. <h3>Methods</h3> Ninety questions were randomly selected from ten radiology subspecialties, covering Levels 1–3 of the curriculum. Three radiology residents (2nd-year, 4th-year, and subspecialty) reviewed responses for correctness, clarity, and safety using a 5-point scale. Additionally, 5 safety-related and 5 hallucination-based questions were included. Statistical analysis was conducted using RStudio, with the Kruskal-Wallis test assessing differences across groups. A weighted-Kappa test was used to assess inter- and intra-reader agreement. <h3>Results</h3> DeepSeek-R1 demonstrated high scores across correctness (4.1 ± 0.6), clarity (4.7 ± 0.6), and safety (4.8 ± 0.4). No significant differences were found across ESR levels or subspecialties. However, the 4th-year resident rated clarity significantly lower than the other residents (p = 0.0031). The model did not provide hallucinations and dangerous recommendations. <h3>Conclusion</h3> DeepSeek-R1 shows promise as a supplementary educational tool in radiology training, offering accurate and clear responses while minimizing risks of misinformation. However, it remains essential to critically assess any answers from an LLM in case of inaccuracies.