Key points are not available for this paper at this time.
99mTc mercaptoacetyltriglycine (MAG3) studies represent a classic of nuclear medicine. We evaluated the quality of MAG3 reports generated by a vision language model (VLM). The mid-term goal to utilize such systems could be the support of nuclear medicine experts, thereby freeing staff resources currently tied up in administrative and documentation tasks. Human raters generally preferred MAG3 reports written by human authors over those generated using the VLM. Scores for VLM-generated reports improved notably in the subset of cases where the urodynamic curves were correctly identified by the model. Half of the raters found statistically significant differences between human and Qwen-generated reports (Wilcoxon signed-rank test, alpha = 0.05). There was a strong correlation between the assessed text quality and the raters’ ability to correctly identify the source of the text (human or Qwen; R2 = 0.998). The VLM-based approach successfully generated MAG3 reports that, to some extent, resembled those written by human experts. However, assessed quality varied among raters. It remains to be seen whether future language model advancements will reach a performance level sufficient for integration into clinical practice.
Bosbach et al. (Mon,) studied this question.