Healthcare
Train medical imaging and monitoring models
For imaging-AI companies, device makers and hospital researchers: train on scans, video and vital signs.
What it's worth
Cut specialist labelling time
- annotator time to segment one image by hand; three annotators typical
- 2–15 minutes
- saved per kidney exam; model-assisted segmentation cut expert segmentation time 90%, one 2025 study
- USD 14.30
- AI-enabled medical devices on the US FDA list; a record 331 authorised in 2025
- 1,400+
Sources: arXiv 2209.15314; Radiology: Artificial Intelligence, 2025; US FDA list of AI-enabled medical devices, 04.03.2026, reported by MedTech Dive.
Use cases
Find the rare case in routine clinical data
| Use case | Raw data | Rare cases to find | Adapt per | Release test |
|---|---|---|---|---|
| Medical imaging | CT, MRI and X-ray in DICOM | Rare pathologies | Scanner, hospital | Sensitivity on a fixed test set |
| Surgical video | Operating-room video | Complications, rare anatomy | Procedure, hospital | Missed events against the last release |
| Patient monitoring | Vital signs, monitor and device logs | Early deterioration, missed alarms | Device, ward | Missed alarms on replayed cases |
| Clinical notes | Dictation and notes in local languages | Rare drug names, mixed languages | Hospital, specialty | Word error rate on a fixed test set |
Medical imaging
- Raw data
- CT, MRI and X-ray in DICOM
- Rare cases to find
- Rare pathologies
- Adapt per
- Scanner, hospital
- Release test
- Sensitivity on a fixed test set
Surgical video
- Raw data
- Operating-room video
- Rare cases to find
- Complications, rare anatomy
- Adapt per
- Procedure, hospital
- Release test
- Missed events against the last release
Patient monitoring
- Raw data
- Vital signs, monitor and device logs
- Rare cases to find
- Early deterioration, missed alarms
- Adapt per
- Device, ward
- Release test
- Missed alarms on replayed cases
Clinical notes
- Raw data
- Dictation and notes in local languages
- Rare cases to find
- Rare drug names, mixed languages
- Adapt per
- Hospital, specialty
- Release test
- Word error rate on a fixed test set
How it works
Raw scans and signals in. Tested releases out.
Melt dataPrepare the data · 01–03
01Ingest
Load DICOM scans, operating-room video, monitor and device logs, and notes; drop corrupt series, catalogue by hospital, scanner and date, and de-identify patient details.
02Mine
Find rare pathologies, complications and early deterioration hidden in routine cases, and drop near-identical routine studies.
03Label
Auto-label findings and outlines with large labelling models inside the platform; your specialists correct a sample of training labels.
Forge intelligenceTrain and test the model · 04–06
04Train
Retrain your shared base model on the full, versioned dataset when your change-control plan allows; every release traces to its exact data.
05Adapt
Fine-tune the base model for each scanner, hospital and device, and tune it again when protocols change and your change-control plan allows.
06Validate
Optimise each release candidate for its device, then test sensitivity and false alarms against the last release on a fixed set.
Reference standards
Keep training labels and test answers apart
Auto-labels train the model; your qualified readers set the answers it is tested against.
- Training labels
- Auto-labels, with a sample your specialists correct. Used for training only.
- Test set
- Every study read in full, disagreements settled and every read logged; then the set is fixed.
- Signed test record
- Each release keeps its dataset and test-set versions, every metric against the last release with its 95% confidence interval, and who signed off and when.
Quality
Regressions stop at the gate
Every release candidate runs the last release's test sets; a drop beyond your margin blocks it until you sign off.
Your team sets the metric and margin for each test set, such as sensitivity for every finding you already detect and false alarms per scan, each with its 95% confidence interval. A margin can read: sensitivity may not drop by more than 3 points.
Healthcare
release v13 → release candidate v14
| Test set | Cases | Last release | Release candidate | Margin | Result |
|---|---|---|---|---|---|
| Nodules 8 mm and larger, sensitivity | 1,240 | 0.92 | 0.93 | −0.03 | Pass |
| Pneumothorax, sensitivity | 410 | 0.95 | 0.95 | −0.03 | Pass |
| Rare: tension pneumothorax, sensitivity | 212 | 0.89 | 0.81 | −0.03 | Fail −0.08 |
| False alarms per scan | 2,000 scans | 0.24 | 0.21 | +0.03 | Pass |
Nodules 8 mm and larger, sensitivity
- Cases
- 1,240
- Last release
- 0.92
- Release candidate
- 0.93
- Margin
- −0.03
- Result
- Pass
Pneumothorax, sensitivity
- Cases
- 410
- Last release
- 0.95
- Release candidate
- 0.95
- Margin
- −0.03
- Result
- Pass
Rare: tension pneumothorax, sensitivity
- Cases
- 212
- Last release
- 0.89
- Release candidate
- 0.81
- Margin
- −0.03
- Result
- Fail −0.08
False alarms per scan
- Cases
- 2,000 scans
- Last release
- 0.24
- Release candidate
- 0.21
- Margin
- +0.03
- Result
- Pass
Release candidate blocked: 1 test set beyond its margin
Interface preview · sample data
Your data
Patient data stays under your control
- Ownership
- Your data-use agreements govern patient data; your labels and trained weights stay yours. Bytemelt has no access to them: your contract says so, and the platform is built so only accounts your team controls can open them.
- De-identification
- Your team de-identifies DICOM headers, burned-in text, names and IDs at ingest and defaces head scans, with the platform's tools and versioned rules you set.
- Location
- Your data is stored and processed only at the partner data centre your contract names, in the Middle East and outside the EU. The contract names the site and country before any data moves, and data leaves the site only when your team moves it. Personal data from the EU moves only under transfer safeguards such as Standard Contractual Clauses, and we give your data-protection officer the site details a transfer impact assessment needs.
- Vendor review
- Before you sign, we answer your vendor-risk questionnaire, and your contract names the data-centre operator and the data-protection terms.
Questions imaging and device teams ask
Which data-protection terms do you sign?
Your contract sets them, agreed with your data-protection officer before any data moves. Your data, above, gives the transfer terms for patient data from the EU.
Does the platform fit our quality system and change-control plan?
It works inside them, not in place of them. Your team exports each release's signed test record, locked once signed, as design-history and change-control evidence.
See the other three industries
Size compute for your clinical data
In one hour, the founder maps your imaging and monitoring models to the six stages and sizes the compute.
contact@bytemelt.ai