Ensemble Architecture
RAD-DINO
Vision Transformer (ViT-B/14)- Pre-trained by Microsoft on radiology images
- Fine-tuned on CheXpert
- Input resolution: 384x384
- ~86M parameters
AUC:
0.8380
DenseNet121
Convolutional Neural Network- Pre-trained on ImageNet
- Fine-tuned on CheXpert
- Input resolution: 320x320
- ~7M parameters
AUC:
0.8470
Ensemble Integration
Weighted Average Mechanism- RAD-DINO weight constraint: 60%
- DenseNet121 weight constraint: 40%
- Optimized entirely on validation set distribution
- Leverages complementary inductive biases
AUC:
0.8523
Validation Performance
| Condition | RAD-DINO | DenseNet121 | Ensemble | Rating |
|---|---|---|---|---|
| Pleural Effusion | 0.892 | 0.901 | 0.908 | Excellent |
| Cardiomegaly | 0.878 | 0.885 | 0.892 | Excellent |
| Edema | 0.810 | 0.840 | 0.845 | Excellent |
| Atelectasis | 0.805 | 0.810 | 0.815 | Excellent |
| Consolidation | 0.805 | 0.799 | 0.801 | Excellent |
| Mean AUC | 0.838 | 0.847 | 0.8523 |
Training & Interpretability
Approach
Progressive unfreezing
Optimizer
AdamW
Loss
BCE with class weights
Method
Grad-CAM (Gradient-weighted Class Activation
Mapping)
Target
Layer
DenseNet DenseBlock4 spatial
convolutions
Dataset Specifications
CheXpert v1.0 Small Stanford ML Group
Total Radiographs: 224,316
Unique Patients: 65,240
Training Split: 153,000
Validation Set: 18,900
Target Labels: 5 (Diagnostic)
Unique Patients: 65,240
Training Split: 153,000
Validation Set: 18,900
Target Labels: 5 (Diagnostic)
REGULATORY NOTICE:
ChestXpert is for educational and research use only. Not
intended or cleared for clinical diagnostics. Consult a qualified radiologist or physician for
medical decisions.