Neurosurgeons favour confidence maps for AI anatomy guidance

When an AI system marks anatomy in a surgical image, a useful explanation may be where the boundary is uncertain. An international survey of 47 neurosurgeons found that certainty scores and maps ranked above explanations based on influential pixels or similar scenes. The study used endoscopic pituitary surgery as its example.[1]

The team, including Danyal Zaman Khan and Tobias Joseph Adams, compared five explanation formats: certainty scores, certainty maps, saliency maps, scene-similarity scores and nearest-neighbour illustrations. Participants ranked them, rated them using a modified Explanation Satisfaction Scale and supplied free-text comments.[1]

Different explanations answer different questions

In computer vision, a saliency map can highlight portions of an image that influenced a prediction. That is useful for researchers checking whether a network attends to unexpected pixels. A scene-similarity comparison may show examples that look like the current frame. Yet an operating surgeon may need something more specific: how sure is the system that a highlighted border is correct, and how sharply does that assurance change across the image?[1]

In endoscopic pituitary surgery, small spatial differences can be important because critical structures are close together. A confidence map aligned to an anatomical boundary is intuitively easier to relate to a decision about dissection. A single confidence score can help indicate overall uncertainty, though it cannot replace spatial detail. The participants tended to regard saliency and similarity explanations as more useful for education than for live clinical navigation.[1]

These preferences concern the explanation display, not the accuracy of the underlying anatomical segmentation. A confidence map can look reassuring even when its probabilities are poorly calibrated. A system that labels comparable predictions as 90% certain should be correct at roughly that rate; otherwise the display may encourage misplaced confidence.[1]

From preference to performance

The next question is whether the preferred format helps surgeons recognise errors and use uncertain information appropriately during a task. The researchers call for prospective evaluation across surgical domains. A cross-sectional survey cannot establish that a display improves operative performance or that the same preference applies to other procedures.[1]

The design implication is specific: explanations should match the information the clinician needs at the moment of use. A saliency map may help someone study how a model behaves; a boundary-confidence map may better communicate where its anatomical guidance becomes uncertain. The survey gives developers a reason to test that distinction.[1]

Report an error or suggest a correction