|
|
--- |
|
|
license: mit |
|
|
tags: |
|
|
- keras |
|
|
- teeth-alignment |
|
|
- dental |
|
|
- healthcare |
|
|
- unsupervised-learning |
|
|
- rlhf |
|
|
- image-classification |
|
|
datasets: |
|
|
- custom |
|
|
library_name: keras |
|
|
language: en |
|
|
pipeline_tag: image-classification |
|
|
--- |
|
|
|
|
|
|
|
|
<h1 align="center">🦷 Teeth Alignment Detection Model</h1> |
|
|
|
|
|
<p align="center"> |
|
|
<img src="https://huggingface.co/VilaVision/dentalmisalignmentdetection/resolve/main/Overbite.jpeg" alt="VilaVision Logo" width="400"/> |
|
|
</p> |
|
|
|
|
|
|
|
|
## 🧠 Overview |
|
|
|
|
|
This Keras model classifies dental images into **aligned** vs. **misaligned** categories. It is designed to aid dental practitioners and orthodontists by analyzing clinical photos or X-rays and detecting signs of malocclusion, crowding, or improper alignment. |
|
|
|
|
|
🧪 **Training Highlights**: |
|
|
- **Unsupervised Learning Phase**: Learns visual features from unlabeled dental image data. |
|
|
- **RLHF (Reinforcement Learning with Human Feedback)**: Fine-tuned using expert-labeled feedback to make the predictions align with real-world diagnoses. |
|
|
|
|
|
> 📌 This model is a research tool and not a substitute for professional dental evaluation. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🏗️ Architecture |
|
|
|
|
|
The model is a Convolutional Neural Network (CNN), built in Keras. It likely includes: |
|
|
|
|
|
- Convolutional layers (Conv2D + ReLU) |
|
|
- MaxPooling or AveragePooling layers |
|
|
- Dense classification layers |
|
|
- Possibly residual connections for stability |
|
|
|
|
|
🖼️ **Input shape**: `(224, 224, 3)` |
|
|
📤 **Output**: Class probabilities (e.g., `[0.8, 0.2]` → "aligned") |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧾 Training Data |
|
|
|
|
|
Though the dataset is not publicly available, it likely contains: |
|
|
|
|
|
- Intraoral or panoramic dental photographs |
|
|
- Images annotated by human experts |
|
|
- Unlabeled data used in the unsupervised phase |
|
|
- Labeled samples used during RLHF fine-tuning |
|
|
|
|
|
The model is inspired by techniques described in [BMC Oral Health, 2022](https://bmcoralhealth.biomedcentral.com/articles/10.1186/s12903-022-02466-x) and [PMC Orthodontic AI](https://pmc.ncbi.nlm.nih.gov/articles/PMC8813223/). |
|
|
|
|
|
--- |
|
|
|
|
|
## 🚀 Usage |
|
|
|
|
|
### 🔧 Install Dependencies |
|
|
|
|
|
```bash |
|
|
pip install tensorflow huggingface_hub |
|
|
``` |
|
|
|
|
|
### 📄 Load and Predict |
|
|
|
|
|
```python |
|
|
from tensorflow import keras |
|
|
from huggingface_hub import hf_hub_download |
|
|
|
|
|
# Download model |
|
|
model_path = hf_hub_download(repo_id="VilaVision/dentalmisalignmentdetection", filename="final_teeth_model.keras") |
|
|
model = keras.models.load_model(model_path) |
|
|
|
|
|
# Preprocess image |
|
|
img = keras.preprocessing.image.load_img("path/to/teeth_image.jpg", target_size=(224, 224)) |
|
|
x = keras.preprocessing.image.img_to_array(img) / 255.0 |
|
|
x = x.reshape((1,) + x.shape) |
|
|
|
|
|
# Predict |
|
|
preds = model.predict(x) |
|
|
print("Raw output:", preds) |
|
|
# Example: preds[0][0] > 0.5 → "misaligned" |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## 📥 Input & 📤 Output |
|
|
|
|
|
| Type | Description | |
|
|
| ------ | ------------------------------------------- | |
|
|
| Input | JPG/PNG image of teeth (224×224), RGB | |
|
|
| Output | Class probabilities for alignment detection | |
|
|
|
|
|
--- |
|
|
|
|
|
## 📈 Performance |
|
|
|
|
|
While no official metrics are available, CNN models for orthodontic imaging tasks have reported: |
|
|
|
|
|
* ~95–98% accuracy ([BMC Oral Health, 2022](https://bmcoralhealth.biomedcentral.com/articles/10.1186/s12903-022-02466-x)) |
|
|
* High F1-scores in clinical benchmarks |
|
|
|
|
|
**Note:** Performance may vary on images that differ from the training distribution. |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚠️ Limitations |
|
|
|
|
|
* Not suitable for diagnostic use without expert supervision |
|
|
* Trained on specific dental image styles — generalization may be limited |
|
|
* May not perform well on low-quality or occluded images |
|
|
* Biases in training data may affect outputs |
|
|
|
|
|
Always consult a licensed orthodontist or dentist before taking action based on model predictions. |
|
|
|
|
|
--- |
|
|
|
|
|
## 📜 License |
|
|
|
|
|
🪪 MIT License – free to use, modify, and distribute. |
|
|
|
|
|
[View on Hugging Face →](https://huggingface.co/AP6621/teeth_alignment_detection_modal) |
|
|
|
|
|
--- |
|
|
|
|
|
## 📚 References |
|
|
|
|
|
* [Deep Learning for Orthodontic Photo Classification – BMC Oral Health](https://bmcoralhealth.biomedcentral.com/articles/10.1186/s12903-022-02466-x) |
|
|
* [AI for Classifying Orthodontic Images – PMC Study](https://pmc.ncbi.nlm.nih.gov/articles/PMC8813223/) |
|
|
* [OpenAI – Learning from Human Feedback (RLHF)](https://openai.com/research/learning-from-human-feedback) |
|
|
|
|
|
--- |
|
|
|
|
|
🧠 *Model built and maintained by [VilaVision](https://github.com/VilaVision)* |
