Thyroid Nodule Malignancy Detector (5-Fold Validated)
This model uses a Dual-Stream Spatial-Frequency Fusion architecture (ConvNeXt-Tiny + FFT Magnitude Spectrum) to classify thyroid nodules in ultrasound images.
📊 5-Fold Cross-Validation Performance
The model was evaluated using a stratified 5-fold cross-validation on the consolidated dataset.
| Fold | Accuracy | AUC | Sensitivity | Specificity |
|---|
| 0 | 0.881728 | 0.94519 | 0.92364 | 0.79386 |
| 1 | 0.866856 | 0.939398 | 0.878661 | 0.842105 |
| 2 | 0.869688 | 0.945065 | 0.899582 | 0.807018 |
| 3 | 0.871013 | 0.934528 | 0.900628 | 0.808791 |
| 4 | 0.858965 | 0.930304 | 0.875523 | 0.824176 |
Summary Statistics:
- Mean AUC: 0.9389 ± 0.0065
- Mean Sensitivity: 0.8956 ± 0.0195
- Mean Specificity: 0.8152 ± 0.0185
🚀 Clinical Application
The weights hosted here (pytorch_model.bin) correspond to Fold 0, which achieved the highest individual AUC of 0.9452.
🛠 Methodology
- Backbone: ConvNeXt-Tiny (Spatial Stream)
- Texture Analysis: FFT Magnitude Spectrum (Frequency Stream)
- Preprocessing: CLAHE
- Loss Function: Focal Loss (α=1, γ=2)
