Vision-Enabled Large Language Models in Mammography: From Benchmarking to Clinical Accountability

Artificial intelligence (AI) has begun to reshape breast imaging, and the accompanying article in this issue illustrates how the next wave may differ from computer-aided detection and task-specific deep learning systems familiar to breast radiologists.1–6 Vision-enabled large language models (LLMs) can ingest images, respond to natural-language prompts, and generate structured descriptions across visual and textual information.7,8 The Breast Imaging Reporting and Data System (BI-RADS) provides standardized terminology and assessment categories.9 Early breast-imaging applications, including diagnostic decision support and BI-RADS category assignment, have shown both promise and important limitations.10–12 The broad applicability that makes LLMs attractive in radiology also makes them difficult to evaluate. The accompanying study, “Benchmark Evaluation of Vision-Enabled Large Language Models in Mammography: A Case-Control Study,” evaluated five contemporary vision-enabled LLMs—ChatGPT-5T, ChatGPT-5I, ChatGPT-4o, Gemini 2.5 Pro, and Gemini 2.5 Flash—in a case-control mammography dataset comprising 100 malignant and 100 control examinations.1 Each model received standard mammographic views and performed lesion detection, laterality assignment, quadrant localization, lesion type classification, and benign-versus-malignant classification.1