A note for people who read for a living

You've heard me complain about PACS for years.
I finally did something about it.

The short version, updated 23 July 2026, after a week where most of it stopped being slideware. Where something is still unfinished I say so, including what is broken today. Forward it to anyone who reads for a living.

The problem, in one sentence

We spend our days inside software built by people who have never had a list of 40 unread studies at 6pm, and it shows: something like 45 minutes a day per radiologist lost to clicking, waiting, re-hanging, and fighting the PACS-to-dictation glue.

Sectra is genuinely good, if you're a 400-bed system with a seven-figure budget. Below that, you're choosing between platforms that were modern in 2012 and cloud products built by people who have never watched someone read. That's the gap.

What Slate.MD is

A reading-first PACS + reporting system for teleradiology groups, imaging centers, and community hospitals, built by a radiologist who still reads daily and tested nightly on my own multi-monitor rig. The whole product is organized around one question: what makes the next study open faster and read cleaner?

What it does

The read

Start with what a modern PACS should simply be and too often isn't: no scrolling glitches, no slowdown across a long shift, a hanging protocol filling every monitor, an interface that is genuinely good to look at. Slate treats those as the floor. On top: STAT mode (sign, and the next study is already hung, the one after already in RAM), every tool binding remappable, and a full-window 2×2 MPR with crosshairs you grab in the middle and drag.

The dictation

You never speak a comma or a period; punctuation and capitalization are inserted for you. Findings supersede the normal they contradict, and the impression is drafted from your findings and the patient's prior reports. Call it thirty seconds a report saved from wrestling the text editor; across a sixty-study shift that is about half an hour back, a real slice of the 45 minutes above.

Start Case Speed ▶ auto-plays when in view
Illustrative replay of the shipping workflow; no audio runs on this page. Grey italic text is the live transcript. Slate decides the organ only after hearing the finding (watch it light up), keeps a continued thought in the same organ, undoes a correction cleanly, and flags a danger pair (medial vs lateral) amber until you click to verify.

That's the transcript half. The same engine then drafts my impression from the findings and the patient's prior reports, streaming as it writes: slip in an ungrounded number and the draft dies mid-sentence; when a restyled sentence can't be proven to still say what I said, my verbatim words stand.

A few things that make the dictation hold up over a real shift:

  • On-site speech engine: your audio never leaves the building. Two passes per utterance, fast partials and an accuracy pass behind them, with a chunk ledger that reports a dropped second instead of quietly swallowing it.
  • It handles the PowerScribe classics: "sigmoid colon" never becomes "sigmoid:", "radiograph" never becomes a paragraph break, and hypo/hyper can never be silently auto-"corrected": words that invert clinical meaning are structurally barred from automatic substitution and flagged for your eyes instead.
  • A per-user voice profile, the thing everyone misses from Dragon: your vocabulary, your corrections, and (with your consent) acoustic adaptation that only ships an update if it measurably beats your current accuracy on your own dictations. You watch your error rate drop month over month.

The report

  • Dictate findings in whatever order you read; they route to the right organ sections deterministically, and when the system isn't sure it leaves them where they are instead of guessing. Last full run: 241 of 241 utterances across twelve scripts landed in the right section.
  • Enforced in the architecture itself: the system can never assert a clinical finding you didn't make. Templates are built from your own prior reports in your own words; suggestions can prompt and style, never author. A complete verbatim record of what you said exists at every instant.
  • The model never sees a number. Measurements are swapped for tokens before any text reaches it and swapped back after, and every restyled sentence is checked against what you said: same findings, same lateralities, same negations, or it doesn't ship. Roughly one section in six comes back verbatim instead of restyled, the check doing its job, mostly in neuro and MSK prose.
  • The impression drafts from your findings and prior reports (yours and, when it helps, a colleague's), streaming as it writes. An ungrounded number kills generation mid-sentence: a fifth of a second on one of my boxes, under half on the other, measured on the live stream. It fails closed: no draft beats a confident wrong one.
  • Chart-aware guardrails: "gallbladder is unremarkable" on a patient with a 2019 cholecystectomy in the chart gets an amber "verify" before it reaches a referrer. Same for left/right vs the order, contrast vs the tech's notes, and "the clinical question was never addressed."
  • A sign-to-send hold your group configures: instant for STAT (the ED is standing there), a couple of minutes for routine, so the oops you catch at Sign never reaches the chart.

The chart

  • A vision-language model reads the priors, the requisition, and the chart, and hands me what matters with a citation back to the source, not a data dump. I don't know anyone else using a VLM for the radiologist's own workflow like this.
  • A PHI-aware context pane puts the story beside the pixels. A comparable prior's stable findings, even across modality (minus CT-only language: "hypodense" means nothing on MR), come up as one-click chips in my own phrasing: "left renal cyst measuring 1.3 cm, unchanged / increased / decreased." A complex or indeterminate lesion on the prior won't let me sign until I've put today's measurement on it; a simple cyst never holds up my signature. Nothing is asserted for me.
  • Named interval procedures drop into the right section, attributed, one click to dismiss: "interval ACDF at C5–C6" lands in spine. A vague "s/p ACDF" with no level stays in the context pane rather than guess.
  • Seamless integration with AI modules: a CT head with a hemorrhage can top your list the moment it lands, so the sick ones surface first. Slate is the surface these tools plug into, not another one to bolt on.

Where it stands

Most of this stopped being slideware in the last few weeks. The reading loop (worklist, hang, scroll, measure, capture, sign) runs end to end on my own three-monitor rig, and the dictation path is up on a GPU turning my real voice into routed findings and a drafted impression. The newest viewer builds and the chart-context work are the last pieces earning their sign-off across full shifts before I call them done. We're lining up design-partner pilots for this fall: paid but design-partner priced, with my cell number and a weekly call.

What I'd love from you

  • If you're in a terad group, imaging center, or hospital department on an aging platform: give me 30 minutes to read a real case in front of you. An actual read, open to sign, not a feature tour.
  • If you're not: tell me the three things about your PACS that make you want to throw a monitor. That list is literally the roadmap; half of what's above started as that kind of complaint.
  • Know a group owner who should see this? Send it along.

Write to me directly

Eli Lim, MD · eli@slate.md
Group inquiries and anything for the company: hello@slate.md