Generative models trained on large-scale data sets have demonstrated an unprecedented ability to generate humanlike text [1] and have performed surprisingly well on untrained tasks (zero-shot learning) [2]. In medical imaging, the applications are manifold, and it has been shown that models can not only draw radiological conclusions [3] but also structure reports [4] and even generate impressions based on the findings given in a report [5] or the image itself [6]. One of the leading obstacles limiting the development of models for generating clinically applicable reports is the lack of evaluation metrics that capture the core aspects of radiological impressions [7,8]. While there are initial studies on the perception of artificial intelligence (AI)–generated text in the general population [9], insights are missing for specialized areas such as medical imaging. Therefore, our study investigated the ability of GPT-4 to generate radiological impressions based on different inputs, focusing on the correlation between radiological assessment of impression quality and common automated evaluation metrics, as well as radiological perception of AI-generated text.

According to the radiological score, the human-written impression was rated highest, although not significantly higher than the text-based impressions (Table 1). A detailed analysis is shown in the supplementary results section in Multimedia Appendix 1. The automated evaluation metrics showed moderate correlations to the radiological score for the image impressions; however, individual scores diverged depending on the input (Figure 1). Correct detection of an impression’s origin (human/AI) varied by input (text: 61/100, 61%; image: 87/100, 87%; radiologist: 87/100, 87%; text and image: 63/100, 63%). For the text input, a homogeneous distribution was found, similar to radiological impressions classified as AI generated (supplementary figure in Multimedia Appendix 1). It was shown that impressions classified as human written were rated significantly higher by the radiologist, with a mean score of 18.11 (SD 1.87) for impressions classified as human written and 13.41 (SD 3.93; P≤.001) for impressions classified as AI generated.

We evaluated the “out-of-the-box” performance of GPT-4 for chest x-ray impression generation based on different inputs. Based on the radiological score, text-based impressions were not significantly lower than the radiological impressions, whereas other inputs were rated significantly lower. Sun et al [5] showed that text-based impressions rated by radiologists were inferior. However, the study did not clarify if the radiological evaluations of the impressions were conducted under blinded conditions. Our work identified radiological bias, as impressions classified as human written received higher ratings. Therefore, without blinding, there is a risk that the inferiority of the AI-generated impressions is due to bias.

For the automated metrics, the impressions based on text and image were rated the closest to the radiological impressions, followed by text-based impressions. For the image-based impressions, there was a significant moderate correlation between the automated metrics and the radiological score; however, for the other inputs, opposite or nonsignificant correlations were found. Automatic metrics that capture relevant aspects of report quality are a prerequisite for successful development and clinical integration. Evaluation metrics, however, can only be as good as the human assessment, which is not free of bias and characterized by false heuristics [9]. Our findings underline this point, as impressions that were classified as human written scored significantly higher in the radiological assessment. Human evaluation is not error-free, but it is the benchmark for the evaluation of generated text.

Radiological heuristics, sources of error, and relevant aspects of radiological quality need to be further investigated, as they are essential for the development of useful model metrics.