Early signal. Human judgment.

Detect risk.
Earlier.

A multimodal deep learning model combining CT scan imaging and RNA-seq genomic data to help identify lung adenocarcinoma staging — supporting earlier clinical decision-making.

Try the Indicator ↓ How it works ↘
Reviewed & Approved by Dr. Tshering Lachenpa, MD
0.618Multimodal AUC
520+RNA-seq patients
3Clinical cohorts
Patient risk indicator
illustrative
CT scan lung visualization
signal map multimodal view
Elevated
87%
!
Further clinical evaluation may be appropriate.
Model status ready v1.0 / demo
02 / the problem
~
1.8m
people die from lung cancer every year worldwide.

Lung cancer is the leading cause of cancer death globally. Detecting potential risk earlier can give healthcare professionals more opportunities to investigate concerning cases.

Lung adenocarcinoma (LUAD) accounts for ~40% of all lung cancer cases. Staging at diagnosis determines treatment and prognosis — yet accurate staging still requires invasive biopsy or extensive imaging workup.
03 / the approach

Turning patient data
into actionable signals.

The model accepts CT, RNA-seq, or both — because in clinical practice not every patient has both. The system adapts to whatever data is available, and the ablation results show exactly when each modality matters.

🧬 Genomic Only
0.636
AUC · best performer

RNA-seq alone captures molecular subtypes — KRAS/STK11 co-mutations and proliferation signatures directly correlate with late-stage disease.

🔬 Multimodal (CT + RNA)
0.618
AUC · flexible input

When both modalities are available. Current linear fusion dilutes genomic signal — future cross-attention with 3D CT may unlock true synergy.

🩻 Imaging Only (CT)
0.576
AUC · no biopsy needed

When no RNA-seq exists. CT is routine standard-of-care — better than nothing when genomic data isn't available.

01
⌁

Analyze

What goes in: A 2D axial CT slice and/or RNA-seq gene expression profile (18,514 genes, HGNC symbols). Upload whichever data you have — CT, RNA-seq, or both.

02
◌

Identify

What the model does: ResNet-18 (RadiologyNET) extracts a 512-dim CT embedding; ComBat + PCA compresses RNA-seq to 480 dims. Available modalities fuse through a linear head.

03
↗

Assist

What comes out: A staging prediction (Stage I/II vs. III/IV) with Grad-CAM saliency and genomic attribution — with a clear indication of which modality drove the result.

Read the output

The indicator surfaces uncertainty to make the next clinical conversation easier to start.

Stage I/IIEarly — fewer spread signals
Stage III/IVLate — consider further evaluation
04 / see it in action

Try the full
model.

Upload a CT scan, RNA-seq file, or both — the actual trained model runs inference and returns a staging prediction with Grad-CAM saliency and genomic attribution.

Open Full Model ↗
Inference runs locally · model weights not uploaded
AI+MD
05 / responsible AI

AI that assists.
Never replaces.

A responsible staging tool should make the next question clearer — not make the decision for the person who owns it.

  • Does not diagnose lung cancer or any other condition.
  • Intended as a decision-support tool, not a clinical endpoint.
  • Requires review by a qualified oncologist or radiologist.
  • Needs external validation before any clinical deployment.
  • Trained on 160 labeled patients — limitations apply.
06 / technology

From signal
to next step.

A transparent pipeline keeps every output legible — from raw patient data through model inference to a structured clinical signal.

01
🩻
Patient Data CT + RNA-seq
→
02
🧬
Preprocessing ComBat + PCA
→
03
🤖
ML Model ResNet-18 + Fusion
→
04
📊
Risk Signal Grad-CAM + XAI
Python 3.12 PyTorch 2.0 ResNet-18 (RadiologyNET) scikit-learn ComBat batch correction PCA (480 dims) Grad-CAM Streamlit 5×5-fold CV Apple MPS TCGA · CPTAC · Stanford

"I think your approach of combining imaging and genetics for staging disease is very interesting, especially that you have been able to achieve a degree of accuracy in your cross validation. I approve this project and wish you all the best in your endeavors."

Dr. Tshering Lachenpa, MD  ·  Reviewed & Approved
The next signal is human

Earlier signals.
Better decisions.

Building technology that helps healthcare professionals see potential staging risk sooner — and act with more information.

Try the Indicator ↓ Not a diagnosis. A decision-support tool.