Janus

Multi-label abdominal disease classification from 3-D CT scans. DINOv3 backbone with four ablation variants: GAP, Masked Attention, Scalar Fusion, and Gated Fusion.

Ablation variants

Variant	Folder	Macro AUROC	Needs seg masks	Needs radiomics
GAP (baseline)	gap/	0.8488	—	—
Masked Attention	masked-attn/	0.8667	✓	—
Scalar Fusion	scalar-fusion/	0.8332	✓	✓
Gated Fusion	gated-fusion/	0.8811	✓	✓

Each folder contains model.safetensors and config.json (labels, image size, input requirements, best macro AUROC).

Quick-start — GAP variant (self-contained)

The GAP variant only needs the CT volume. No segmentation or radiomics required.

pip install torch transformers safetensors nibabel scipy huggingface_hub
python inference.py ct.nii.gz --variant gap --top 10

Example output:

Disease                                    Prob
──────────────────────────────────────────────────
  Atherosclerosis                         0.821  ████████████████
  Renal Cyst                              0.714  ██████████████
  Hepatic Steatosis                       0.431  ████████
  ...

Full pipeline (masked-attn / gated-fusion / scalar-fusion)

These variants require organ segmentation masks (and radiomics features for fusion models). Use the full Janus preprocessing pipeline:

Source code: https://github.com/lavsendahal/janus (Repository is currently private — request access if needed.)

git clone https://github.com/lavsendahal/janus
cd janus && pip install -e .
python -m janus.inference --variant gated-fusion --ct ct.nii.gz

Model details

• Backbone: DINOv3 ViT-S/16 (tri-planar 2.5-D slicing of 3-D CT)
• Task: 30-label multi-label binary classification (abdominal CT findings)
• Training dataset: MERLIN abdominal CT dataset
• Input: [B, 1, D, H, W] CT volume, HU clipped to [-1000, 1000] and normalised to [0, 1]

Diseases (30 labels)

Hepatomegaly, Splenomegaly, Cardiomegaly, Prostatomegaly, Hepatic Steatosis, Osteopenia, Gallstones, AAA, Aortic Valve Calcification, Coronary Calcification, Atherosclerosis, Thrombosis, Bowel Obstruction, Appendicitis, Hiatal Hernia, Submucosal Edema, Free Air, Biliary Ductal Dilation, Absent Gallbladder, Pancreatic Atrophy, Hydronephrosis, Renal Cyst, Renal Hypodensities, Pleural Effusion, Atelectasis, Ascites, Anasarca, Metastatic Disease, Lymphadenopathy, Fracture.

Citation

If you use this model, please cite the Janus paper (forthcoming).