---
language: en
license: apache-2.0
tags:
- medical
- thyroid
- ultrasound
- convnext
- explainable-ai
metrics:
- auc: 0.9389
- accuracy: 0.8697
---
# 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)