Imagine a doctor spending hours meticulously checking maps of a patient's organs, drawn by another AI, to prepare for precise radiation therapy. It's a critical but time-consuming task. A new study tested whether a different kind of AI—a large language model—could act as a first-line assistant to help with this quality check.
The research used 20 male pelvic CT scans. The AI assistant, called LAQUA, was asked to evaluate the quality of organ outlines created by other auto-contouring software. When compared to the judgments of two expert radiation oncologists, the AI's ratings showed moderate to strong agreement. In a specific test, it was very good at correctly identifying acceptable contours for the rectum and at spotting problematic ones for the left femoral head. The AI also provided written explanations for its ratings, which human reviewers found were generally well-aligned with the actual contour quality.
However, this is a very early, small-scale look at the idea. The study used only 20 scans from one part of the body (the male pelvis), so we don't know if it would work for other cancers or for female patients. The researchers themselves caution that their method might miss some poor-quality contours, and some performance metrics had wide confidence intervals, meaning we need more data to be sure. No safety issues were reported, but this was a technical analysis of scans, not a treatment trial. The core finding is simply that this approach is feasible enough to explore further as a potential time-saver for busy clinicians.
