Author:
Ghosn Youssef,Sardouk Omar El,Jabbour Yara,Jrad Manal,Kamareddine Mohammed Hussein,Abbas Nada,Saade Charbel,Ghanem Alain Abi
Abstract
AbstractObjectiveTo evaluate the performance of two versions of ChatGPT, GPT4 and GPT3.5, on the Final FRCR (Part A) also referred to as FRCR Part 2A radiology exam. The primary objective is to assess whether these large language models (LLMs) can effectively answer radiology test questions while providing accurate explanations for the answers.MethodsThe evaluation involves a total of 281 multiple choice questions, combining the 41 FRCR sample questions found on The Royal Collage of Radiologists website and 240 questions from a supplementary test bank. Both GPT4 and GPT3.5 were given the 281 questions with the answer choices, and their responses were assessed for correctness and accuracy of the explanations provided. The 41 FRCR sample questions difficulty was ranked into “low order” and “high order” questions. A significance level of p<0.05 was used.ResultsGPT4 demonstrated significant improvement over GPT3.5 in answering the 281 questions, achieving 76.5% correct answers compared to 52.7%, respectively (p<0.001). GPT4 demonstrated significant improvement over GPT3.5 in providing accurate explanations for the 41 FRCR sample questions, with an accuracy of 65.9% and 31.7% respectively (p=0.002). The difficulty of the question did not significantly affect the models’ performances.ConclusionThe findings of this study demonstrate a significant improvement in the performance of GPT4 compared to GPT3.5 on FRCR style examination. However, the accuracy of the provided explanations might limit the models’ reliability as learning tools.Advances in KnowledgeThe study indirectly explores the potential of LLMs to contribute to the diagnostic accuracy and efficiency of medical imaging while raising questions about the current LLMs limitations in providing reliable explanations for radiology related questions hindering its uses for learning and in clinical practice.HighlightsChatGPT4 passed an FRCR part 2A style exam while ChatGPT3.5 did not.ChatGPT4 showed significantly higher correctness of answers and accuracy of explanations.No significant difference in performance was observed between “high order” and “lower order” questions.Explanation accuracy was lower than correct answers rate limiting the Models’ reliability aslearning tools.
Publisher
Cold Spring Harbor Laboratory
Reference10 articles.
1. Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models;PLOS Digit Heal,2023
2. Bhayana R , Krishna S , Bleakney RR. Performance of ChatGPT on a Radiology Boardstyle Examination: Insights into Current Strengths and Limitations. Radiology. 2023;307(5).
3. The Royal Collage of Radiologists. Final FRCR Part A – Guidance for Candidates. :1–32. Available from: https://www.rcr.ac.uk/sites/default/files/how_to_approach_cr2a_candidate_guidance.pdf
4. The Royal Collage of Radiologists. Final FRCR Part B Examination - Purpose of Assessment Statement. Available from: https://www.rcr.ac.uk/sites/default/files/cr2b_purpose_of_assessment_statement.pdf