For the reading room
For the AI lab
Two sides of one radiology department.
Radyssey keeps the cases you teach from and keeps an eye on the models you build. One company, two products, the same rules for patient data.
RadBase, step by step
- 01Upload the DICOM study and its report.
- 02De-identify headers and burned-in text inside a locked quarantine.
- 03Review: a consultant confirms no identifiers remain and checks the AI's tags.
- 04Library: search by sign, system or diagnosis and open the full study.
- 05Learn from cited reference passages beside the images.
RadShell, release by release
- NowRunBoard: watch long jobs, then launch and stop them.
- R1Projects & registry · Data & labels
- R3Scan review
- R4Models & training
- R5Testing
- R6Deploy & monitor
radbase.co
Great cases shouldn't disappear.
The best teaching cases live in phone galleries, WhatsApp groups and a consultant's personal folder. When that resident rotates out, the cases go with them. RadBase keeps them: one shared, de-identified library the whole department learns from.
- Quarantine
The raw upload goes into a locked quarantine that only the de-identification step can read.
- De-identify
Names, MRNs and dates are removed from image headers and from text burned into the pixels, including local name and MRN formats.
- Sign-off
Nothing is published until a consultant confirms no identifiers remain. The raw copy is deleted within 72 hours.
- Tags
AI reads the report and suggests tags across ten axes: modality, anatomy, diagnosis, imaging signs, scoring systems. A human confirms each one.
- Library
Filter by sign, system or diagnosis, and open the full study, not a screenshot.
- Cited
Every case comes with learning material from trusted references, with citations you can open and check.
Your data stays yours. Your best cases reach everyone.
Each hospital keeps its own RadBase server and its own data. When it chooses, it shares de-identified cases to a common pool, so a resident in one city can learn from a rare case seen in another.
radshell
Long jobs, shown only as far as known.
Inference, segmentation and training runs take hours on lab machines nobody is watching. RunBoard, RadShell's first screen, shows what each one is doing without anyone having to log in and check.
- Progress read from the job's own log: done of total, rate and ETA.
- A plain state: Running, Waiting, Done, Failed, or "Status unknown since …" when a log goes silent.
- Each machine's GPU, load, memory, temperature and last-seen time.
- A read-only live log tail, and open items from the project's checklists.
Watching is read-only: RunBoard never writes to, moves or stops anything on a watched machine. Launching and stopping runs is the next step.
> tail [job].log [log line] [log line]
At a glance
| Question | RadBase | RadShell |
|---|---|---|
| Who it's for | Residents and consultants | AI lab researchers |
| What it keeps | De-identified teaching cases, tagged and searchable | Job progress, machine health and log tails |
| Where it runs | On-premise, one server per hospital | On the lab's own machines and network |
| Status | Coming soon | In development |
The rules both products follow
AI suggests. Humans decide.
Every AI output is a teaching aid that a person confirms. Neither product is a diagnostic tool.
Patient data stays in the hospital.
Raw images are quarantined and de-identified before anyone sees them, and RadBase can run entirely on-premise.
Show only what is known.
No guessed progress bars. When a log goes quiet or a machine drops off, RadShell says so.