Data infrastructure for medical imaging

Over 90% cost reduction.
One standardized format.

Compress and standardize whole-slide images at ingest. No scanner, storage infrastructure, or viewer changes needed.

Accepting pilot institutions

Pipeline Active
logit v0.1.0
SourceSizeStatusOutput
Ingested0 MB
Output0 MB
Ratio—
Files0

Over 100x

median compression

visually lossless quality

MS-SSIM ≥ 0.99

standardized output

OME-TIFF/DICOM

drop-in deployment

no scanner or viewer changes

How it works

From scan to searchable.

Every file is compressed, embedded, and standardized in a single pass.

NDPIbreast_core_bx_04245.ndpi

2.4 GB → 24 MB100xlogit pipeline

01 INGEST

Listens for any scanner format

DICOM, SVS, NDPI, MRXS, BIF. No conversion required at source.

40+ FORMATS

02 EMBED

Generate tile-level vector embeddings

Each tile is embedded in the same pass. This enables semantic search across your entire archive.

1024-DIM VECTORS

03 COMPRESS

Visually lossless compression

Tiles and compresses each slide at ≥ 0.99 MS-SSIM quality. Median 90%+ size reduction.

≥ 0.99 MS-SSIM

04 STANDARDIZE

Convert to OME-TIFF or DICOM

Standardized output. Pyramidal tiling for fast viewing at any zoom level.

OME-TIFF / DICOM

05 STORE

Write to existing storage

No change to your existing storage or viewer infrastructure. Drop-in deployment.

S3 · AZURE · ON-PREM

Waiting for file...

≥0.99 MS-SSIM · visually lossless

Built with input from pathologists and researchers at Penn Medicine

Performance

Compression without compromise.

Decode times measured on standard hospital hardware.

Decode Benchmarksingle core · single tile · compressed output

Tile decode time vs. one frame at 60 fps

256×256 tile1.5ms

full bar = 16ms (1 frame @ 60fps)

512×512 tile4ms

No perceptible delay. Existing viewers work unchanged.

* Measured fetching and decoding one tile on a single CPU core. Multi-core systems decode tiles in parallel, with typical hospital workstations handling hundreds of tiles simultaneously.

Cost Estimator

See what you're spending.

Storage costs compound as data accumulates. Plug in your numbers.

$168.9K

saved over 1 year

Reduce storage costsRequest a pilot

Cost Estimator

100x compression

Slides / year

Avg file size

Storage cost

$/GB/mo

>90% cost reduction.

* median compression ratio exceeding 100x, measured across 32,500 uncompressed tiles at 512×512 with MS-SSIM ≥ 0.99.

Platform

What the ingest point unlocks.

Every tile that flows through the compression pipeline is automatically embedded and indexed. That index powers everything below.

Semantic Search

Find any tile across your entire archive by example. Get results in milliseconds. No annotation required.

Logit StudioSearch

LibraryDatasetsPipelines

Search across 2.4M indexed tissue tiles

Try:

Dataset Creation

Define datasets by example. Provide 5 reference tiles, retrieve 50,000 curated training samples in seconds. Weeks of annotation become minutes.

Dataset Builderlogit studio

Source

5 reference tiles · invasive ductal carcinoma · breast

Query

cosine similarity > 0.89
filter: H&E stain only

Querying index...

Results

0 tiles retrieved from 0 slides
12 institutions · 4 scanner types
Query time: 4.2s

training_set_idc_v3

47,832 tiles · 1024-dim embeddings

Proceed to training→

Model Development

Train models directly on curated datasets. Your data never leaves your infrastructure.

Model Traininglogit studio

Cell [1]

running

▶ Run

1from logit.studio import Dataset, Trainer

3dataset = Dataset.load('training_set_idc_v3')

4model = Trainer(

5 arch='resnet50',

6 embeddings='UNI',

7 dataset=dataset,

8 epochs=20

10model.train()

Output

Epoch 1/20

loss: 0.4821val_acc: 0.847

░░░░░░░░░░✓

Epoch 5/20

loss: 0.2103val_acc: 0.921

░░░░░░░░░░✓

Epoch 10/20

loss: 0.0847val_acc: 0.963

░░░░░░░░░░✓

Epoch 15/20

loss: 0.0412val_acc: 0.978

░░░░░░░░░░✓

Epoch 20/20

loss: 0.0289val_acc: 0.984

░░░░░░░░░░✓

Training...

Validation

Held-out slides: 500

Accuracy: 0.981Sensitivity: 0.974Specificity: 0.988AUC: 0.996

Ready for deployment

Deploy model→

Deploy & Monitor

Deploy models into production. Detect distribution drift automatically when scanners, stains, or tissue types shift.

Deployment

tumor_classifier_v2

logit studio

STATUS

deployed

14 days active

SLIDES

processed

AVG CONF

0.94

mean confidence

THROUGHPUT

~920

slides/day

Drift Monitor

Distribution baseline: training_set_idc_v3

Day 1–10····················nominal

Day 11····················▲alert

Day 12–14····················nominal

▲Alert·Day 11·2 slides flagged

Embedding distance > 2σ from training distribution
Scanner: Leica SCN400 · Stain: atypical PAS protocol

Action: slides queued for pathologist review

Your platform for AI deployment in healthcare.

Compression reduces storage costs. Standardization enables interoperability. Embeddings power search, dataset creation, and model deployment.

Over 90% cost reduction.One standardized format.

From scan to searchable.

Compression without compromise.

See what you're spending.

What the ingest point unlocks.

We build data infrastructure for medical imaging.

Over 90% cost reduction.
One standardized format.