5CNet · In production inside Bionic Vision
A vast study.
A small signal.
5CNet is 5C Network's neural architecture for sparse anomaly detection in high-dimensional medical imaging. It runs in production today and powers Bionic Vision.
The operation 5CNet runs inside
- scans a day
- 20,000+
- scans a day
- hospitals
- 2,000+
- hospitals
- radiologists
- 400+
- radiologists
- cities
- 300+
- cities
- studies reported
- 20M+
- studies reported
- clinical images
- 4B
- clinical images
The idea
Most of a study is normal. The diagnosis lives in the rest.
A single CT or MRI holds an enormous amount of image data. The evidence that decides the diagnosis, a small nodule or a thin fracture line, usually occupies a tiny fraction of it. Everything else is normal anatomy.
Radiology is sparse anomaly detection followed by context-conditioned inference: find the few regions that do not look as they should, then decide what they mean given the whole patient and the whole study. 5CNet is built around that fact rather than around a list of diseases.
Sparsity, written down
ρ = |S| / N ≪ 1
- S
- the regions that carry diagnostic evidence
- N
- everything in the study
An outlier is not automatically pathology. Unusual is a reason to look closer, not a diagnosis.
Normal is judged per anatomy. What is normal for a lung is not normal for a liver.
A small finding must not become a small priority.
The 5C Block
One block. Five steps. Every scale.
5CNet is built from one repeated unit. The same five steps run on the whole study, then on each series, slice and region, down to a single patch.
- 1
Context
Understand the whole study first: what was scanned, how, and why.
- 2
Compare
Predict what normal anatomy should look like, and measure how far each region deviates from it.
- 3
Concentrate
Spend extra compute only on the regions that are unexpected. The rest has already been seen.
- 4
Confirm
Look for support across adjacent slices, other views or sequences, and prior studies.
- 5
Characterize
Turn the evidence that survives into a structured finding, with its uncertainty stated.
Repeated at every scale
Dual path
See everything. Spend compute selectively.
A global path reads the whole study, so nothing is skipped and every finding is interpreted in context. Alongside it, a specialist path looks closely at only the few regions the global path marks as unexpected. The two meet before anything is called a finding.
Confirm
A candidate is a question, not an answer.
Before 5CNet calls something a finding, it looks for the same evidence elsewhere, the way a radiologist scrolls to the next slice or pulls up last year's scan. Evidence that holds up is kept. Evidence that does not is set aside.
Candidate region
Unexpected against normal anatomy
-
Adjacent slices
Does it continue through neighbouring slices the way real anatomy would?
-
Other views and sequences
Is it visible on another projection or MR sequence?
-
Prior studies
Is it new, changed, or stable against earlier scans?
Confirmed evidence
Carried forward to Characterize
Why an architecture, not a list of detectors
More detectors, more false alarms.
Much of radiology AI is a stack of single-disease classifiers, each scoring the whole study for one condition. That design has two problems for a hospital reading every study, every day.
Averaging can erase a small finding
When a whole study is compressed into one score, a signal that covers a tiny fraction of it can be diluted by everything that is normal. The smaller the finding, the easier it is to lose.
False alarms multiply
If each of K independent detectors has a false-alarm rate f, the chance that at least one fires on a normal study is:
P(any false alarm) = 1 − (1 − f)K
| Detectors (K) | 1 | 10 | 50 | 100 |
|---|---|---|---|---|
| At f = 1% | 1% | 9.6% | 39.5% | 63.4% |
Arithmetic illustration only. Not a measurement of any product.
| A stack of single-disease detectors | 5CNet |
|---|---|
| One whole-study score per disease | Evidence kept as a set of regions |
| Every detector fires independently | Evidence confirmed and reconciled before any disease call |
| Unusual is treated as abnormal | An outlier is not automatically pathology |
| One threshold for the whole study | Normal judged per anatomy |
From evidence to signed report
Find what matters. Complete the report. Learn from the work.
5CNet does not output a list of disease scores. It outputs a structured finding set: what was found, where, why, and how certain, with localisation and key images. Everything downstream is built from that.
-
01 · 5CNet
Structured finding set
What, where, why, how certain, with key images.
-
02 · Draft
Report drafted from the findings
Built from the finding set, not from a blank page.
-
03 · Bionic LM
QC agents check the report
Contradictions, omissions, the clinical question answered.
-
04 · Radiologist
Reviews and signs
Every report. The diagnosis is theirs.
The model family
One architecture, four modalities.
-
5CNet-XR
Radiographs
-
5CNet-CT
Computed tomography
-
5CNet-MR
Magnetic resonance
-
5CNet-Mammo
Mammography
What it learns from
Knowing what normal looks like is the starting point. The rest comes from the daily work of reporting.
- Normal studies, so it knows what normal looks like
- Study-, finding- and evidence-level labels
- Radiologist edits to drafted reports
- Escalations
- Discrepancy reviews
Where it sits
Underneath Bionic.
Hospitals do not run 5CNet as a separate tool. It sits beneath the Bionic products, inside the 5C Radiology OS that routes, reads and reports every study.
5C Radiology OS
Bionic
Architecture
5CNet
Sparse anomaly detection for medical imaging. Powers Bionic Vision.
Questions about 5CNet
What is 5CNet?
5CNet is 5C Network's neural architecture for sparse anomaly detection in high-dimensional medical imaging. It treats a scan the way a radiologist does: most of the study is normal, and the evidence that decides the diagnosis sits in a small part of it. 5CNet finds that evidence, confirms it, and turns it into a structured finding.
Is 5CNet used in clinical practice today?
Yes, 5CNet runs in production at 5C Network today and powers Bionic Vision, the AI inside the reporting workflow 5C Network runs for 2,000+ hospitals at 20,000+ scans a day. Every report is still signed by a radiologist.
How is 5CNet different from single-disease radiology AI?
5C Network designed 5CNet to treat evidence as a set that is confirmed and reconciled before any disease call, instead of stacking one detector per disease. Each independent detector adds its own false alarms, and averaging a whole study into one score can erase a small finding. In 5CNet an outlier is not automatically pathology, and normal is judged per anatomy.
What is the 5C Block?
The 5C Block is the five-step unit that 5C Network's 5CNet repeats at every scale, from the whole study down to a single patch: Context, Compare, Concentrate, Confirm and Characterize. It understands the study, measures deviation from expected normal anatomy, focuses compute on unexpected regions, seeks confirming evidence, and produces a structured finding with uncertainty.
Does 5CNet replace the radiologist?
No, 5CNet does not replace the radiologist at 5C Network. 5CNet produces a structured finding set, the report is drafted from it, Bionic LM's QC agents check it, and a radiologist reviews and signs every report. 5CNet is built so a small finding does not become a small priority.
Which modalities does 5CNet cover?
5C Network builds 5CNet as a family of models: 5CNet-XR for radiographs, 5CNet-CT, 5CNet-MR and 5CNet-Mammo. Each learns what normal looks like from normal studies, and learns from labels, radiologist edits, escalations and discrepancy reviews.
A small finding must not become a small priority.
See how Bionic, running on 5CNet, fits into your reporting, with a radiologist signing every report.
