Impression policy
The impression is the part of a report that is read, so GigHz Precision AI treats it differently from the findings. The findings are what you dictated, placed where they belong. The impression is composed, ordered and filtered by rules that are the same for every report and that you can read on this page.
Your own impression always wins
If you dictate an impression ("Impression: one, acute appendicitis…"), it is used verbatim. It is renumbered and image citations are removed; nothing else is touched, ordered or filtered. Everything below applies only when the impression is composed for you.
Composition
A composer model is given the dictated findings, the dictated indication and limitation if any, and the study type, with these instructions:
- Determine the clinical question: the indication if dictated, otherwise inferred from the limitation or the study (a CT pulmonary angiogram is answering pulmonary embolism).
- Answer that question first, explicitly, even when the answer is negative.
- Then list the remaining findings by clinical severity.
- A time-critical finding always becomes item one, above the indication.
- Vascular studies must characterise the vessels.
- Measurements and laterality are copied exactly; image and series numbers never appear.
- Normals and pertinent negatives are excluded; every distinct abnormal finding is included; similar lesions are summarised as a group with a size range.
Deterministic ordering
The composer's lines are then ranked by a priority table that does not depend on the model. Active bleeding, solid organ injury, aortic catastrophe, tension physiology and intracranial haemorrhage sit at the top; torsion, bowel ischaemia and perforation, cord compression, acute stroke and limb ischaemia next; then abscess, mass, fracture, aneurysm, inflammatory processes, obstruction, thrombosis, effusions, nodules requiring follow-up; incidental and degenerative findings last. A term inside a negated clause ("no active extravasation") never ranks a line. "Chronic", "stable", "unchanged" and similar words demote a line out of the emergency tiers, so a chronic pulmonary embolism is not treated as a massive one.
Simple cysts and small lymph nodes are gathered into one prose sentence rather than numbered lines; a cyst with features (septated, nodular, enhancing, complex) keeps its own line.
What is kept out
The rule, from the radiologists who shaped it: if an indication is given, the impression should answer it, and otherwise mention only what is urgent or needs follow-up. So the following never appear in a composed impression unless the same line carries an escalating term (complex, numerous, fractured, loosened, obstructing, new, increased, suspicious…) or the finding is itself the answer to the indication:
- simple atelectasis;
- unchanged degenerative or postoperative change of the bones, including intact hardware;
- simple cysts (hepatic, renal, punctate);
- atherosclerotic and other vascular calcifications;
- diverticulosis without diverticulitis;
- a collapsed or under-distended bladder;
- artifact and limitation statements.
A second tier, including uncomplicated hernias, lipomas, gallstones without cholecystitis, hepatic steatosis and reactive nodes, stays in the impression only when no indication was given or nothing else answers it. If every line is removed, the impression reads "No acute abnormality."
Completeness check
After composition, a backstop compares the impression against the dictated findings. Any abnormal finding missing from the impression is appended, through the same sanitiser and the same suppression rules, so the backstop can neither drop a finding nor reintroduce an incidental one.
Per-template control
Through Make it mine a template can set its own impression behaviour: differentials when a finding is indeterminate, a standard follow-up recommendation, or neither; and ordering by priority (default), as composed, or emergencies-only.
Why this is deterministic on purpose
The ordering and suppression tables are configuration files that ship with the application and are checked by the automated test suite on every release. The behaviour of the impression should not depend on a model's mood: a splenic laceration with active extravasation is item one every single time.
This page is generated from documentation that ships with the application (docs v1.2.2). Version numbers, prices and trial lengths on this page are read from the running system, not typed by hand.