MSU Abdominal Ultrasound / CC BY 4.0
bv-abdus-v1 · BlackVoxel
Organ recognition on ultrasound.
Classification of ten structures in abdominal ultrasound frames.
What it does
The model identifies the structure in the frame for routing and quality control. The classes refer to organs, not diseases.
- Model
- bv-abdus-v1
- Ownership
- BlackVoxel
- Base
- ResNet-18 / ImageNet transfer / 10 classes
- Data
- MSU Abdominal Ultrasound / CC BY 4.0
Demo
Image, attention map and draft report.
The screen uses previously processed results. Inference is not run on this page.
Kidney · 100%
B-MODEW/L · AUTO
STORED MODEL OUTPUTGRAD-CAM
Training and evaluation
Data and setup.
ResNet-18 / ImageNet transfer / 10 classes
Task-specific labels
independent test
ResNet-18 fine-tuned on 4,109 labeled frames. Training uses radiologist 1 data and testing uses the entire independent radiologist 2 set.
Accuracy 0.820, macro-AUROC 0.978 and macro-F1 0.732 on the independent 1,334-frame test set.
Test results
Measured performance.
macro-AUROC
independent testaccuracy
ten classesmacro-F1
n=1,334The test uses data from a second radiologist group. Recall for the portal vein class was 0.091.
Limitations
Scope of this result.
- 01
Recognizes the organ; it does not detect stones, aneurysms or focal lesions.
- 02
Single frame, not the examination sweep.
- 03
Rare classes have unstable estimates and low recall.
Next model
MG / ROI