SonoBelief weights
Frozen components of SonoBelief. Code: github.com/daryataratynova/sonobelief.
Trained on TotalSegmentator v2.0.1 (302 train studies), selected on validation (38 studies), seed 0. The models take simulated label-map views rendered from CT, not ultrasound images.
| Folder | Component |
|---|---|
image_encoder/ |
E_img: view β 128-d embedding |
liver_shape/ |
E_L: liver mask β shape code z_L (32-d) |
observation_model/ |
D_obs: (z_L, probe pose) β view embedding, plus ΟΒ² |
flow_planes/ |
placement flow pre-trained on random views |
flow/ |
placement flow fine-tuned on sweeps (used) |
pose_head/ |
g_pose: target score from pose only |
geometry_head/ |
g_geom: target score from pose + inferred GB geometry (SonoBelief) |
exact_geometry_head/ |
g_GB: target score from pose + true GB geometry (analysis only) |
python scripts/weights.py download
Not a medical device.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support