Clinician Evaluation Protocol
Draft formative-research protocol / 31 October 2026 / version 0.2
Intended Use for This Release
Explore public or otherwise authorized, deidentified research CT data, compare supplied anatomy masks with multiplanar images, and collect structured feedback about visualization and workflow. Do not use this release to diagnose, select a resection, determine margins, guide instruments or make patient-care decisions. Imported volumes have no AI segmentation.
Reviewers and Data
Invite a small representative formative panel, for example cardiothoracic surgeons and a thoracic radiologist, under an agreed institutional research/data-governance process. A small usability panel is not a statistically powered clinical validation study. Agree the protocol, reference standards and acceptance criteria with qualified clinical and regulatory advisers before any clinical performance study.
The bundled cases s0310 and s0223 are anatomy teaching examples from TotalSegmentator. They do not represent the full range of surgical pathology, contrast phases, institutions or scanners. Derived DICOM fixtures are generated from public NIfTI data; they are not original scanner acquisitions. The synthetic phantom has known coordinates and voxel values and tests software geometry, not medical accuracy.
Formative Tasks
- Confirm right/left, anterior/posterior and superior/inferior orientation against an independent reference viewer.
- Locate corresponding anatomy in axial, coronal, sagittal and 3D displays.
- Inspect every supplied label, flag implausible borders and truncation, and record disagreements. A research-review flag means inspected, not accurate or approved.
- Confirm scale and intensity warnings are visible. The distance tool must remain disabled for unspecified spatial units.
- Import supported deidentified research data and verify that unsupported or inconsistent series are rejected without changing the previous study.
- Export and restore an encrypted review; compare source case identity, coordinates, notes and warnings.
- Record task completion, misunderstandings, discrepancies and the severity of observed usability concerns.
Feedback and Acceptance
The application records five operator-entered concern ratings, including a not-assessable option. Scores are not a calibrated confidence probability, prediction of effectiveness, clinical sign-off or endorsement. Do not translate them into a percentage of production readiness. Preserve disagreement and negative feedback.
For a later clinical study, preregister the intended clinical question and independently reviewed reference data; address selection bias, scanner/protocol diversity, segmentation errors, boundary failures, inter-reader variability, clinical risk and the human-software team. Determine sample size and acceptance criteria with appropriate statistical expertise. Do not reuse training data as independent clinical testing.
Stop Conditions
- Unexpected laterality, geometry, scale, intensity or source-study mismatch.
- Missing slices, unsupported encoding, inconsistent Frame of Reference or mixed series.
- Unexplained data loss, misleading review status, missing warnings or incomplete exports.
- Identifiable data, uncertain authorization, or proposed use in actual patient care.
Record the issue, retain only appropriately governed evidence, and suspend the affected workflow until investigated. Geometry checks and passing automated tests must not overrule a clinician-observed discrepancy.