Breast ultrasound classifier — DenseNet121 v2b (research only)
⚠️ Not a medical device. Research model trained on public datasets; must not be used for diagnosis.
Five DenseNet121 fold models (v2b: mask-guided training + aspect-ratio crop augmentation) for benign vs malignant classification of breast ultrasound lesions. Code, data audit and full evaluation: https://github.com/wajidali99/breast-ultrasound-classification
Decision rule
Each model outputs P(malignant). A model votes malignant if P ≥ 0.335 (threshold frozen on BUSI out-of-fold predictions for ≥90% sensitivity). The label is the majority of the 5 votes.
Performance (ensemble AUC)
| Data | AUC |
|---|---|
| BUSI 5-fold CV (pooled out-of-fold) | 0.940 |
| BUSI locked test set (105 images) | 0.970 |
| BUS-BRA external (1,875 images, Brazil) | 0.776 (95% CI 0.749–0.804) |
Limitations
Large performance drop on an external hospital (AUC 0.97 → 0.78): results do not transfer reliably across sites.
At the BUSI threshold, external sensitivity was 0.70 (target 0.90); site-specific calibration would be required.
Trained on ~530 BUSI images from a single centre; images without a lesion (normal) are out of scope.
The uploaded weights are a deterministic rebuild of the evaluated v2b models; four folds reproduce the original validation AUC exactly and fold 4 differs by 0.002. Reported BUS-BRA results come from the original v2 run.
- Downloads last month
- -