When an AI system marks surgical anatomy, a surgeon needs to know where its outline is dependable and where it is uncertain. An international survey of 47 participants working in neurosurgery found that confidence scores and certainty maps were the most useful of five explanation formats they were shown. Their preferences favoured information tied directly to judging a boundary in the operative image.[1]
The researchers investigated the interpretability of AI decision-support systems for surgical anatomy recognition, using endoscopic pituitary surgery as a high-risk example. Such operations can involve narrow approaches near critical structures, where an incorrect identification could have consequences beyond a mislabeled photograph. The team did not deploy an automated surgical robot or measure whether an AI overlay saved patients from injury. It asked participants which kinds of explanation they found useful when an AI system identifies anatomy.[1]
Five explanation approaches were considered. The participants ranked certainty scores, certainty maps, saliency maps, scene-similarity scores and nearest-neighbour illustrations. They also rated the approaches using a modified Explanation Satisfaction Scale and gave free-text responses. These are distinct ways to communicate model behaviour: a certainty score gives a numerical indication of confidence, a certainty map distributes it across the image, and a saliency map points to portions of an image influential in a model's output.[1]
Matching the explanation to the decision
The respondents consistently preferred certainty-based explanations. Certainty scores were valued for communicating confidence, while certainty maps were valued for showing anatomical boundaries and local areas of doubt. Both more closely resembled the decisions surgeons already make while judging tissue relationships. Saliency maps and similarity-based techniques were generally regarded as less clinically relevant, though some respondents saw value in education or training. The message is about matching the information to a specific task rather than crowning one universal form of explainable AI.[1]
Five explanations shown to participants
| Format | Information conveyed |
|---|---|
| Certainty score | A numerical indication of model confidence |
| Certainty map | Where confidence varies across the anatomical image |
| Saliency map | Image regions influential in the model output |
| Scene similarity score | How closely the scene resembles reference material |
| Nearest-neighbour illustration | A comparable example from the reference set |
As an illustration, an image might contain one clearly visible vessel and another boundary obscured by lighting or tissue deformation. A saliency display can highlight what influenced the model; a certainty map can show where the predicted boundary is less secure. They answer different questions, which helps explain why the surgeons did not value them equally.[1]
The language of confidence also conceals a statistical problem. A model can report high certainty and still be wrong. Calibration—the relationship between its scores and observed error rates—may differ between hospitals, imaging devices and patient anatomy. A preferred display therefore needs to be evaluated together with the reliability of the model behind it.[1]
The survey used endoscopic pituitary surgery as its example and measured participants’ rankings, satisfaction ratings and written feedback. It establishes an interface preference. The next empirical question is whether that preference helps surgeons accept correct guidance and reject incorrect guidance during actual tasks.[1]
From a preferred display to a calibrated one
A prospective task could measure that directly: show clinicians both correct and incorrect overlays, then examine whether each explanation helps them distinguish the two. That would connect the survey’s interface preferences to the decisions that matter during surgery.[1]
The paper discloses that Hani Joseph Marcus is employed by and holds shares in Panda Surgical. Danail Stoyanov holds shares in Panda Surgical and Odin Vision and is employed by TouchSurgery, Medtronic. Those interests are relevant context for a study of how surgical AI should present its outputs.[1]
The design lesson is specific: an explanation should serve the operator’s decision. Showing influential pixels may help developers investigate a network, while showing a weak boundary may help a surgeon decide where to inspect more closely. Both can be legitimate explanations without being interchangeable.[1]
The survey gives developers a direction grounded in the people who would use the system. For these surgeons, the most valuable explanation was information about where the model’s visual judgement was secure enough to consider and where it needed a second look.[1]
Sources & further reading
- Exploring clinician interpretability of AI decision support systems for surgical anatomy recognition Scientific Reports · 24 Oct 2026 · Accessed 26 Oct 2026