Pathology image analysis

A pathology image model belongs in sign-out only when it returns a quality flag: the slide ID plus a visual region a pathologist can open. That is the entire useful output. It is not a diagnosis, not a grade, and not a report.

If the slide is ordinary, the output stays empty. Empty is success. The model does not invent a grade. The pathologist still signs.

Keep the vendors in one class. PathAI, Paige, and the EHR or imaging panes that may display the same kind of overlay (Epic, Microsoft) all propose something a person must confirm. None of them replace the signature. If a product in this class cannot name a slide and a region, it is not producing the quality outcome this page describes.

The same rule shows up in radiology AI triage and flagging: a locatable mark for a human read, not a substitute for the signed interpretation.

Open the whole-slide image in the viewer you already use to sign. Do not adjudicate a flag from a thumbnail, a PDF export, or a chat snippet. Confirm that the slide ID on the flag matches the label on the glass and the ID in the laboratory information system. If they disagree, drop the flag. Do not guess which slide the model probably meant.

Go to the cited region at diagnostic magnification. A cite is a place: coordinates, a box, a polygon, or a viewer link that lands on that field. A sentence with no location is not a cite.

If the accession has multiple blocks, levels, or immunohistochemistry slides, load the specific slide named in the flag. A result that says invasive carcinoma or high-grade with no slide ID and no field is a grade with no region. That output cannot be confirmed, so it cannot enter the report. Ignore it or send it back. Do not type it in for now.

A flag raised on hematoxylin and eosin does not travel onto the matching immunostain. If you need the IHC, open that slide yourself. Frozen sections are the same rule with less time: load the named frozen, open the region, and do not let a permanent-section overlay stand in for the freeze.

Loading is also when you see artifact. Folded edge, ink, air-dried smear, coverslip debris, and scanning blur all fool region detectors. You still look at the tissue. The flag only told you where the model looked.

A usable flag has three parts.

1. Slide ID, in the same form you would use to pull the glass.
2. A visual region you can pan to in the viewer.
3. A reason you can confirm or reject at the microscope, written as morphology, not as a finished grade.

Stop after those three. Do not let the pipeline attach a Gleason score, a Grade Group, a mitotic count, or a suggested line from diagnosis code suggestion. Coding follows a signed diagnosis. The image model does not get to skip the tissue.

If the viewer paints a heatmap instead of a box, the heatmap is not a cite until you can land on a field. A wash over the entire core is a suspicion, not a region. Tighten it to a navigable box or discard it. If one slide carries several flags, each flag needs its own region. A list of adjectives under a single slide ID is not several cites.

One illustrative pass, not a case study. A prostate needle-core accession arrives with twelve slides. The model returns one flag: slide 7 of 12, a box over a glandular focus at mid-core, reason "crowded small glands with amphophilic cytoplasm, confirm or reject." You open slide 7, land on the box, and decide whether those glands are carcinoma, adenosis, or crush artifact. You write the diagnosis you would have written from the glass. The flag named a place to look. It did not grade the case.

If that same accession had returned Grade Group 3 with no slide and no box, you would have a grade with no region. That is unusable. The same standard applies to clinical photography classification: a label without a locatable image region is not a finding you can stand behind.

Most slides on a tray should produce nothing. If the model cannot cite a slide ID and a region, it returns empty. Do not backfill. Do not invent malignancy so the AI log is not blank.

A silent tray is not a benign diagnosis. It is a machine state: no proposed region. You still review the tissue. If a workflow badge paints no findings in green, read that badge as the model did not propose a region, never as benign.

When a flag does appear, keep it bound to that slide and that region. Do not promote it to the patient (this person has cancer) and do not copy it onto unlabeled levels. Case-level language is how a region flag turns into a fake report.

Questions about trial eligibility are a different workflow, clinical trial matching. That work starts from a signed diagnosis and criteria, not from an unsigned box.

The pathologist signs. The model does not. Residents, fellows, and pathologist assistants may open the region and draft. They do not convert a flag into a signed line. If Epic, a Microsoft imaging pane, or any other desktop shows model text beside the report editor, that text is a viewer aid. It is not the report. Do not paste the flag into the diagnostic line and click sign.

Treating the flag as the report is the failure mode that looks most like efficiency. The interface often renders model output as a finished sentence: invasive carcinoma, Grade Group 2. That sentence is not yours until you have seen the field and written it. If you did not open the region, you did not confirm it. If you opened it and disagree, you remove the flag from the clinical record and write what you saw.

Inventing malignancy is the failure mode that fills silence. It happens when the tray looks too clean and someone adds a positive call so the tool appears to have worked, or when a weak visual cue is promoted to carcinoma because the model pane is open. If you cannot point at glands, nuclei, or architecture on the slide, you do not have malignancy. Absence of a flag is not a prompt to create one.

The working sequence is short. Load the slide the flag names. Keep a flag only when it cites slide ID and region. Leave ordinary slides empty. Sign what you saw.

If a vendor in this class (PathAI, Paige, or an Epic- or Microsoft-hosted viewer surfacing the same kind of overlay) cannot give slide ID plus region, or if it emits a grade as the primary object, keep that output out of the sign-out path. The quality outcome is a cited flag or silence. Anything else will get transcribed into a report it does not belong in.