PYCAD SSR Engine · OHIF integration

Works with OHIF 3Works with OHIF 2No fork, no rebuild

Open multi-gigabyte studies in OHIF in seconds

The PYCAD SSR Engine plugs into the OHIF viewer you already run. Large CT and tomosynthesis series stay on the server; the browser receives only the images on screen. Your PACS, your workflow and your OHIF stay the same.

2.3 sinstead of 188 sMPR on screen for a 5 GB CT in OHIF 3
9 MBinstead of 5.1 GBData sent to the browser for that same study
0.9 GBinstead of 10.9 GBPeak browser memory, OHIF 2 MPR on a 5 GB CT
97%instead of 3%Of slices shown while scrolling breast tomosynthesis in OHIF 2

The problem

Studies outgrew the browser

Photon-counting CT, thin-slice protocols and breast tomosynthesis produce series of several gigabytes. A browser-based viewer has to download and decode all of it before MPR or 3D can start.

On a good connection that means minutes of waiting. On a home connection, MPR may never finish, the browser can use more than 10 GB of memory, and readers on laptops feel it first.

Our approach

Render on the server, keep OHIF

The engine reads the series from your PACS once, renders slices, MPR planes and 3D on a server with a GPU, and streams just the pixels each viewport needs.

Window/level, zoom, pan and measurements still run in the browser on the images it receives, so they feel instant, and pixel values match stock OHIF, HU and ROI statistics included.

How it fits into OHIF

Added like a plugin, not a rewrite

  1. One line in the OHIF config

    A connector script is loaded through OHIF’s documented extension mechanism in app-config.js. No fork of OHIF and no rebuild.

  2. Your PACS stays the source of truth

    The study list, metadata, reports and segmentations keep coming from your PACS over DICOMweb. The engine only renders, and it never writes to the PACS.

  3. Only large series are routed

    Before downloading anything, the connector checks a series’ size. From 200 MB (configurable) it goes through the engine; smaller series stay on stock OHIF. MPR and 3D always use the engine.

  4. Smooth scrolling, full detail at rest

    While scrolling, the viewer receives fast previews sized to the pane; when you stop, the full-resolution slice arrives about 150 ms later.

Results

Measured in stock OHIF 3 and OHIF 2

Each test opens the same study in the same viewer, once with the engine and once without. Times are until the view is on screen; shorter bars are better.

OHIF 3 · MPR and 3D

MPR, 5 GB CT

OHIF + PYCAD engine
2.3 s
Stock OHIF
188 s

MPR, 2.5 GB CT

OHIF + PYCAD engine
1.6 s
Stock OHIF
93 s

3D, 2.5 GB CT

OHIF + PYCAD engine
5.8 s
Stock OHIF
83 s

MPR, 320 MB CT

OHIF + PYCAD engine
2.5 s
Stock OHIF
13.7 s

On a home connection (50 Mbit/s), stock OHIF 3 did not show MPR within 5 minutes on the 2.5 GB and 5 GB CTs; with the engine it took 2.1 s and 3.0 s.

OHIF 2 · scrolling and MPR

Slices shown while scrolling tomosynthesis (higher is better)

OHIF + PYCAD engine
97%
Stock OHIF
3%

Slices shown scrolling a 2.5 GB CT on a home connection

OHIF + PYCAD engine
93%
Stock OHIF
0%

Peak browser memory, MPR on a 5 GB CT (lower is better)

OHIF + PYCAD engine
0.9 GB
Stock OHIF
10.9 GB

MPR open on a laptop, 611-slice CT

OHIF + PYCAD engine
0.6 s
Stock OHIF
26 s

Across all large series, the engine sent 10 to 60 times less data to the browser: 6 to 60 MB instead of 73 MB to 5.1 GB.

Feature parity

Your readers keep the tools they know

We tested the engine against OHIF’s own features, on clinical-style public datasets: CT, MR, dental CBCT, PET/CT and breast tomosynthesis.

MPR and oblique planes

Crosshairs that move and rotate to any oblique plane, with the exact full-resolution plane at rest.

Thick slabs

MIP, MinIP and average slabs, matching classic OHIF on the PET/CT MIP view.

3D rendering

OHIF’s 25 volume presets render on the server and look the same as vtk.js from the same camera.

Measurements

Length, angle, ROIs with HU statistics and probe, with values identical to stock OHIF. In OHIF 2 the engine adds measurements in MPR, which stock OHIF 2 lacks.

Segmentation

Labelmap editing, brushes and DICOM SEG loading, with the same voxel counts as stock OHIF. RTSTRUCT and PET/CT fusion work too.

Reports

Structured reports are saved to your PACS and reopen as usual; the engine itself stays read-only.

Server capacity

Sized by CPU cores, with one GPU for 3D

One engine process serves every reader, sharing loaded studies between sessions. We simulated radiologists scrolling, reformatting and rotating 3D without pauses, a heavier load than real reading.

About 20 readersscrolling smoothly per 8-vCPU server under this heavy synthetic load; 40 on 16 vCPU
1 GPUcovered 3D for 100 simulated users; most of the work is on the CPU
0 errorsin about 222,000 requests, from 5 to 100 simulated users

Smooth means 95% of scroll frames under 60 ms. As a rule of thumb, plan about 0.4 vCPU per actively scrolling reader. Memory grows with the number of different large studies open at the same time.

Being precise

Where the engine doesn’t change much

  • Small series on a good connection. Stock OHIF is already fast, which is why series under 200 MB stay on the stock path.
  • Scrolling stacks in OHIF 3 on a good connection. First-image times are the same as stock; the gain is 5 to 60 times less data and about half the browser memory.
  • Slow mobile connections. At 15 Mbit/s with high latency, neither path scrolls very large CTs smoothly, although the engine shows the first image sooner.
  • Measurements on reformatted views can’t be stored in a DICOM structured report, because there is no acquired image to reference.

Before production we are adding an automatic fallback to stock OHIF if the engine can’t be reached. Ultrasound, microscopy and 4D studies have not been tested yet.

How we tested

Stock viewers, real network, public data

ViewersStock OHIF 3.13 and the last OHIF 2 release (4.12.51), each with and without the engine connector
ServerOne cloud server with 8 vCPU, 52 GB RAM and an NVIDIA T4 GPU (16 vCPU for one capacity run), with a test PACS on the same machine
ClientA Windows laptop running Chrome over the real internet, plus simulated home (50 Mbit/s) and mobile (15 Mbit/s) connections
StudiesA public kidney CT (320 MB), simulated photon-counting CTs of 2.5 GB and 5 GB built from it, and a public breast tomosynthesis study
Runs142 automated browser runs and a capacity test from 5 to 100 simulated users, November 2026. Results on the direct connection are medians of 3 runs.

Running OHIF with large studies?

Tell us your OHIF version, your PACS and the size of your largest series. We’ll tell you what the engine can change for your readers.

Book a call

The PYCAD SSR Engine is not cleared as a medical device. Results were measured with public or synthetic data and depend on the network, the server and the studies. OHIF is an open-source project; PYCAD is not affiliated with it.

We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.

Copyright © 2026 PYCAD. All Rights Reserved.