Radiotherapy Safety Review and Quality Assurance with a Vision-Language Agent

Errors in radiation treatment planning can have serious consequences, while detecting them often requires reviewing information from many clinical documents and images. We developed a local vision-language agent that reviews uncurated records for potentially serious safety concerns. The agent was built on Qwen3.6-35B-A3B-FP8. From a real-world archive of 39,528 folders of patient-specific data from radiation treatment courses, we randomly selected 3,000 records that had undergone standard clinician-based quality assurance and clinical approval. The agent successfully analyzed 2,967 of 3,000 records (98.9%); the remaining 33 records (1.1%) could not be processed because of missing or unreadable source files, context-length overflow, or document-verification errors. The agent flagged 36 of the 2,967 analyzed records (1.2%) for potential errors. Across analyzed records, 22/2,967 (0.7%) had a majority-supported alert and 14/2,967 (0.5%) had an alert that was not majority-supported. Three expert reviewers reviewed every alert. By majority vote, 22 alerts (61.1%) were supported, including three (8.3%) judged clinically important or critical. Separately, 23 alerts (63.9%) had possible or definite workflow value. All inference remained within hospital infrastructure. These findings support prospective evaluation of local vision-language agents as a second layer of clinical safety review.