LIDC-IDRI / TCIA / CC BY 3.0
bv-chestct-v1 · BlackVoxel
Lung nodule characterization.
Binary characterization of a crop centered on a lung nodule.
What it does
The radiologist marks the nodule location. The model classifies the crop and displays the score for morphological review.
- Model
- bv-chestct-v1
- Ownership
- BlackVoxel
- Base
- ResNet-18 / ImageNet transfer / 128 px crop
- Data
- LIDC-IDRI / TCIA / CC BY 3.0
Demo
Image, attention map and draft report.
The screen uses previously processed results. Inference is not run on this page.
Nodule with suspicious pattern · 100%
AXIALW/L · AUTO
STORED MODEL OUTPUTGRAD-CAM
Training and evaluation
Data and setup.
ResNet-18 / ImageNet transfer / 128 px crop
Task-specific labels
patient-split test
ResNet-18 fine-tuned on LIDC-IDRI nodule crops. The random sample was split by patient; the test set contains 97 nodules.
AUROC 0.931, sensitivity 0.750 and specificity 0.967 on the test set, with 36 nodules labeled as malignant-suspicious.
Test results
Measured performance.
AUROC
patient-split testsensitivity
0.5 thresholdspecificity
same thresholdThe AUROC of 0.931 refers to characterization of centered nodules. The evaluation does not include detection, screening or volumetric reading.
Limitations
Scope of this result.
- 01
The nodule must first be localized by the radiologist.
- 02
2D slice and crop, not the complete volume.
- 03
Radiological suspicion is not pathological confirmation.
Next model
US / ABD