Inside LungAI CADx

Three Models. One Probability.
Every Nodule.

A 3D Vision Transformer nodule encoder, a Sybil-derived whole-scan encoder, and a longitudinal decoder, working together to score malignancy risk from a single LDCT.

Request a Technical Brief See Clinical Evidence
100M
Parameter
Vision Transformer
100K
CT scans, 23K
patients trained
0.96
AUROC on NLST
vs Lung-RADS 0.88
~2min
Inference, embedded
in CADe viewer
Live View ⟶ CADe + LungAI CADx
aView · LCS
Summary
Findings
LungAI
Axial · Slice 147/320 · W:1500 L:-600 · GE LightSpeed VCT
LungAIv2.4.1
Patient
Sample_002
M · 67 · 38 PY
Study
LDCT Chest
10/10/2025
Acquisition
320 slices
1.25 mm
0.32 mGy
Prior
04/15/2024
✓ registered
Eligibility
✓ USPSTF
✓ 20+ PY
✓ Age 50-80
Axial · Nodule 1
W: 1500
L: -600
Slice 147/320 · 1.25 mm
RUL · juxta-pleural
+2.1 mm vs prior · VDT 187d
14.3 mm N1
LungAI Score
Nodule 1 · RUL Solid
87/100
High Risk
03070100
Lung-RADS v2022
Category
Very Suspicious
4X
Nodule Detail
TypeSolid
Major axis14.3 mm
Volume1042 mm³
Density+38 HU
MarginSpiculated
LocationRUL · juxta
Longitudinal
Prior size12.2 mm
Δ Diameter+2.1 mm ↑
VDT187 d
Workflow
CADe + LungAI CADx
Read time
~2 min
Output
Category + 0–100 score
Billable
Standard fees + CPT 0721T · $650
Synthetic example. LungAI CADx scores every detected nodule from 0 to 100, segments it, measures it in 3D, and auto-compares to the prior scan, all inside the existing CADe viewer.
The Technology

Where Rule-Based Scoring Reaches Its Limit

Lung-RADS is the clinical standard, and it should be. LungAI doesn't replace it. It fills the gap between what a size threshold can see and what a 100-million-parameter model can.

Lung-RADS v2022
Rule-based
Size thresholds only
Diameter buckets determine the risk category
No patient context
Age, sex, and smoking history are not factored in
No texture or tissue analysis
Density gradients and surrounding tissue are ignored
Manual longitudinal comparison
Radiologist must reconcile prior scans by eye
Categorical output only
Labels 1, 2, 3, 4A, 4B, 4X with no probability estimate
Limited by what rules can see
Oatmeal Health's LungAI
Deep learning
Full 3D image analysis
Operates on the entire DICOM voxel tensor, not just diameter
Patient-specific inputs
Age, sex, and pack-years smoked are explicit model inputs
100M-parameter vision transformer
Captures morphology, texture, and tissue context automatically
Automatic longitudinal tracking
Registers prior scan DICOM and compares nodules across timepoints
Continuous 0–100 score
The physician sees gradations, not just categories
Augments Lung-RADS ↑
LungAI is designed to augment Lung-RADS, not replace it. The reading physician uses both: the Lung-RADS category for guideline-concordant documentation, and the LungAI score as an additional data point when adjudicating borderline cases.
How It Works

A Malignancy Score That Goes Beyond the Diameter

LungAI is a CADx medical device that reads the full DICOM image, not just the nodule's size, and returns a continuous 0–100 malignancy probability inside your existing workflow.

What LungAI Is
Computer-Aided Diagnosis (CADx) as a Medical Device

LungAI accepts a low-dose CT scan as input and outputs a continuous malignancy score from 0 to 100 for each pre-selected pulmonary nodule. A higher score means a higher probability that the nodule is cancerous.

Unlike Lung-RADS, which uses categorical size thresholds, LungAI uses a 100-million-parameter vision transformer that operates on the full DICOM image. It incorporates nodule morphology, surrounding tissue, patient age, sex, and smoking history, and when available, a prior scan for longitudinal comparison.

The score is delivered inside your existing PACS workflow as a PDF report alongside the CT, with no separate platform to log in to.

0–100
Continuous malignancy score
per nodule, not just a category
How LungAI Works
1
LDCT scan acquired
Patient undergoes standard low-dose CT screening
2
CADe flags nodules
Partner device marks nodule centroids on the scan
3
LungAI (CADx) scores each nodule
100M-parameter vision transformer analyzes the full DICOM tensor, considering nodule type (solid, part-solid, GGO), prior scan change, and patient demographics
4
Report delivered in CADe/PACS
PDF with 0–100 malignancy score per nodule appears inside the radiologist's/pulmonologist's existing viewer
5
Physician decides
Radiologist/Pulmonologist uses LungAI score alongside Lung-RADS to determine follow-up
Model Inputs

Three Dimensions Lung-RADS Cannot See

LungAI incorporates nodule type, change over time, and patient demographics as explicit model inputs. Each one shifts the malignancy probability in ways a size threshold alone never could.

Nodule type classification

LungAI distinguishes between nodule types and weights malignancy probability differently for each. Ground-glass and part-solid nodules carry different risk profiles than solid nodules of the same diameter.

Solid Part-solid Ground-glass (GGO)
Longitudinal change detection

When a prior LDCT is available, LungAI automatically registers the nodule across scans and incorporates interval change into the malignancy score. No manual comparison required.

Prior scan registration Interval growth Automatic
Patient-specific risk weighting

Age, sex, and pack-years smoked are explicit model inputs. A 6mm nodule in a 72-year-old with 40 pack-years is scored differently than the same nodule in a 51-year-old with 20 pack-years.

Age Sex Pack-years smoked
The Output

A Number That Tells You What a Category Cannot

Every nodule gets a 0–100 malignancy score delivered directly in PACS. No new login, no separate platform. Just a probability that sits next to the image you're already reading.

LungAI Malignancy Score: 0–100
94
Example
One score per nodule, delivered in PACS
Each nodule receives a 0–100 score. A higher number means a higher probability of malignancy. The score appears alongside the nodule image in the CADe software or PDF report, with an estimated probability curve based on retrospective study data.
How to read the scale
0–30
Low risk
31–60
Intermediate
61–100
High risk
The report also provides a Lung-RADS combined probability table: when the reading physician has assigned a Lung-RADS category, LungAI shows the estimated malignancy probability for that specific score and category combination, giving the physician a richer picture than either input alone.
Important
LungAI is a decision support tool. It does not replace the reading physician and makes no treatment recommendation.
The score is one additional data point. The radiologist reviews it alongside Lung-RADS, nodule characteristics, and clinical history before determining follow-up.
Why a continuous score matters
Lung-RADS assigns a nodule to a category. LungAI assigns it a probability. A nodule that sits at the boundary of 4A and 4B looks the same under categorical scoring. With a continuous 0–100 score, the physician can see whether it sits at 61 or 98, and act accordingly. That gradient is exactly where clinical judgment lives.
Clinical Performance

The Numbers Behind The Score

In our internal study on the NLST held-out dataset, LungAI demonstrated strong sensitivity and specificity, an AUROC of 0.96, and read times under 3 minutes when paired with a CADe platform.

Validated on a held-out dataset, pre-FDA

In our internal study, LungAI demonstrated 94% sensitivity and approximately 88% specificity, an AUROC of 0.96, and a read time of 1.5 to 3 minutes when paired with a CADe platform.

Learn more about the Data
NLST held-out set, retrospective, pre-FDA clearance. Results may differ in prospective clinical use.
0.96
AUROC
NLST held-out, retrospective, pre-FDA
94%
Sensitivity
at 82% specificity
88%
Specificity
at 91% sensitivity
<3 min
Read time
Up to 80% faster than manual reads
Metric
LungAI + Lung-RADS
Lung-RADS alone
False positive rate
Reduced
18–40%
Patient personalization
Age, sex, smoking, prior scan
None
Longitudinal tracking
Automatic
Manual
Output format
PDF in PACS, CPT 0721T billable
Radiologist interpretation only
Implementation

Fits Into What You Already Have, Billable From Day One

LungAI plugs into your existing CADe and PACS environment with no new workstation or workflow changes, and every screened patient generates a new CPT 0721T billing event.

How It Plugs In
Native CADe and PACS integration
LungAI sits inside the workflow the radiologist is already in. No new login, no separate viewer, no disruption.
Works with any CADe or PACS system
No new workstation or workflow changes. Plugs alongside whatever CADe the radiologist is already using.
Score shown on the CADe screen, per nodule
The LungAI malignancy score appears directly next to each nodule's characteristics inside the CADe viewer.
PDF report also delivered in PACS
Per-nodule score, nodule image, and probability curve for documentation and overread billing.
GE, Philips, Siemens. Cloud or on-prem
HIPAA-secure Google Cloud or Docker on-premise with REST API endpoint.
How It Gets Paid
Reimbursable at ~$650 per scan
A new billable event on every screened patient using a dedicated AI CPT code, with no additional imaging or patient visit required.
CPT 0721T: ~$650 per scan
AI-specific quantitative image analysis code, covered by Medicare and major commercial insurers.
Billed by the reading radiologist or pulmonologist
The physician bills directly. No added administrative burden on the imaging center.
Companion code CPT 0722T also available
For additional AI analysis components where applicable.
New revenue on scans already being performed
Monetizes existing CT volume with no additional patient visit or imaging required.
The Full Picture

LungAI Scores the Nodule. LungIQ Finds the Patient.

High-quality screening starts before the scan. LungIQ uses large language models to identify USPSTF-eligible patients who have never been offered screening, so every scanner slot is filled with the right patient.

Upstream Product
LungIQ: Finding the Right Patients
LungAI scores nodules after the scan. LungIQ is what gets the right patients to the scanner in the first place. It uses large language models to mine EHR data and identify USPSTF-eligible patients who have never been offered screening, even when their smoking history is buried in clinical notes.
0.97
AUROC
LLM-powered EHR mining
Parses clinical notes, prescriptions, diagnoses, and procedure codes to surface 20+ pack-year smoking history even when undocumented in structured fields.
Ranked candidate list
Each patient receives a P(candidate) score. Outreach teams work top-down, maximizing yield per call. 94% precision in the top 100 patients contacted at Stigler FQHC.
Works with any EHR
Compatible with Epic, ECW, Athena, Azara, and NextGen. Validated on a live cohort of 35,000 patients with no manual data cleaning required.
For Imaging Centers and Health Systems

Ready to See LungAI in Your Workflow?

Whether you want a live demo in your CADe viewer or want to build a full lung screening program, the next step takes less than 30 minutes.

For Radiologists and Department Heads
See LungAI in your workflow

Request a live demo showing how LungAI appears inside your CADe viewer, what the 0–100 score looks like per nodule, and how the PDF report is delivered into PACS. Less than 30 minutes.

Request a workflow demo
For Imaging Program Leaders and CMOs
Build a lung screening program

Talk to our partnerships team about how Oatmeal Health drives scan referrals to your center, integrates LungAI into your CADe workflow, and enables your radiologists to bill CPT 0721T on eligible scans.

Talk to our team