AI models and limitations
GigHz Precision AI uses large language models for three narrow jobs and deterministic rules for everything that must not vary. Knowing where the line is tells you where to look when you review a report.
Where models are used
| Job | Model role | Deterministic layer around it |
|---|---|---|
| Transcription | Speech to text, streamed | Silence-filler gate (a lone "okay" during a pause is discarded), dictation corrections, spoken punctuation |
| Placement | For each sentence, choose the template line and the operation | Fixed set of allowed operations; a dictated finding always replaces a default; hand-edited lines are protected; unplaceable text is kept verbatim |
| Impression composition | Write numbered lines from the findings and indication | Priority ordering, incidental suppression, image-citation removal, de-duplication, completeness backstop |
| Study routing fallback | Propose a template when phrase matching fails | May only choose from the real library |
| Reference lookups | Answer protocol and appropriateness questions | Scoped to the curated libraries |
Clinical scoring systems (LI-RADS, Bosniak, Fleischner and the rest) are not model output: their features are parsed by rules and their categories come from the published tables, implemented once each. The model is only ever asked to place the sentence; it is never asked to categorise.
Model selection
The application uses current OpenAI models through the API, with a stronger model automatically selected for complex or high-risk cases (an active scoring-system case, or dictation that flags as complex). The specific models are configured by the administrator and change as better ones become available; the documentation does not pin them because the behaviour that matters is governed by the deterministic layers above, which are tested on every release.
Known failure modes, and what catches them
- A sentence lands on the wrong line or is parked. Visible on screen as you dictate; parked text is flagged. A named-vessel finding that the model parks is redirected to its vessel line by rule.
- A sentence is split by a pause. Fragments are re-joined before placement; the engine decides sentence boundaries from the words.
- The composer omits a finding. The completeness backstop re-adds any dictated abnormal finding missing from the impression.
- The composer includes an incidental finding. The suppression rules remove it, at every door, after the backstop.
- The transcriber invents a filler during silence. Filler-only utterances are discarded before they reach the transcript.
- The placement service is down. Findings are parked with a warning and no impression is composed from unplaced text.
- A wrong number. Measurements are copied, never computed; the test corpus checks that dictated numbers appear verbatim. Review every measurement anyway.
What the model cannot do
It does not see the images. Every finding in a report was dictated by you; the software organises, it does not observe. It does not know the patient's history beyond the indication you state. It does not check your findings against each other for clinical consistency beyond the scoring systems.
Regulatory position
GigHz Precision AI, as sold, is a dictation-based drafting and decision-support tool: it organises text the radiologist dictates and applies published classification rules to that text. It does not analyse medical images. Image-analysis features that generate findings from pixels are a regulated device category in the United States and are not part of the product; any such capability is confined to internal research by the developer and is not available to users.
The reviewer's responsibilities
The Terms of Service (version PRECISION-TOS-2026-08-31-v0.9) require that every report be reviewed by the radiologist before use. In practice that means: read every line of the findings against what you intended; check every measurement and laterality; read the impression as the referrer will; confirm or dismiss any scoring panel; and treat Additional findings as a to-do list. The application is built to make that review fast. It is not built to make it optional.
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.