ResNet50 fine-tuned to classify 6 intraoral conditions from a single photo — benchmarked against 3 other architectures, evaluated on a held-out test set.

🏆 94.77% Accuracy · 0.9411 F1-Score (macro)

Model Description

This repository hosts the winning checkpoint from a 4-way benchmark of image classifiers trained to recognize intraoral conditions:

Calculus · Caries · Gingivitis · Ulcers · Tooth Discoloration · Hypodontia

Four architectures were trained under identical conditions (same data split, same augmentation, same evaluation protocol) and compared on a test set none of them saw during training:

Rank	Model	Params (trainable)	Test Accuracy	Test F1 (macro)
🥇	ResNet50 (this checkpoint)	23,520,326	94.77%	0.9411
🥈	DenseNet121	6,960,006	94.51%	0.9351
🥉	EfficientNet-B0	4,015,234	94.17%	0.9335
4	ScratchCNN (no pretraining)	11,179,590	83.45%	0.8236

ResNet50 (ImageNet-pretrained, fine-tuned in two stages — freeze then unfreeze) came out on top and is the model served by Gradio.py and packaged as checkpoints/best_model.pth.

Intended Use & Limitations

This model is a research and educational tool for preliminary visual screening. It is not a certified diagnostic device and must not be used to replace examination by a licensed dentist or physician. Performance depends on image quality and lighting similar to the training data, and may not generalize to conditions or populations outside the training distribution.

Files in this Repository

Path	Description
`checkpoints/best_model.pth`	Final ResNet50 checkpoint — a dict with `state_dict`, `model_name`, `class_names`, and `test_f1`.
`checkpoints/`	Also contains the individually saved weights for the other 3 models trained in the same run.
`notebooks/`	The full training notebook — data prep, transforms, model definitions, training loop, evaluation.
`outputs/`	`models_comparison.csv`, per-model loss/accuracy curves, and confusion matrices.
`Gradio.py`	Standalone web demo — upload an image, get the predicted class, confidence, and full probability breakdown.

How to Use

Load directly with `huggingface_hub`

python

from huggingface_hub import hf_hub_download
import torch

weights_path = hf_hub_download(
    repo_id="nsr51324/Oral_Diseases_Image_Classification",
    filename="checkpoints/best_model.pth"
)

checkpoint = torch.load(weights_path, map_location="cpu")
class_names = checkpoint["class_names"]

Rebuild the model and run inference

python

import torch.nn as nn
from torchvision.models import resnet50
from torchvision import transforms
from PIL import Image

model = resnet50(weights=None)
model.fc = nn.Sequential(
    nn.Dropout(0.3),
    nn.Linear(model.fc.in_features, len(class_names))
)
model.load_state_dict(checkpoint["state_dict"])
model.eval()

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

image = Image.open("sample.jpg").convert("RGB")
tensor = transform(image).unsqueeze(0)

with torch.no_grad():
    probs = torch.softmax(model(tensor), dim=1)[0]

pred = class_names[probs.argmax().item()]
print(f"{pred}: {probs.max().item()*100:.2f}%")

Run the interactive demo

bash

pip install torch torchvision gradio pillow huggingface_hub
python Gradio.py

Point MODEL_PATH at the top of Gradio.py to your local copy of checkpoints/best_model.pth.

Training Data

Trained on nsr51324/Oral_Diseases, sourced from the Oral Diseases dataset on Kaggle (salmansajid05), split 80/10/10 (train/val/test) with stratification to preserve class balance across all three sets.

Training Procedure

Image size: 224×224 · Batch size: 32 · Epochs: up to 30 (early stopping)
Two-stage fine-tuning: backbone frozen for 5 epochs (head-only training, lr=1e-3), then fully unfrozen for fine-tuning at lr=1e-5
Regularization: dropout (0.4), weight decay (1e-4), label smoothing (0.1), early stopping on validation loss
Augmentation (train split only): random resized crop, horizontal flip, rotation, color jitter, random erasing

Evaluation

Full classification report, confusion matrix, and training curves for all 4 models are in outputs/. Summary metrics are in outputs/models_comparison.csv.

Disclaimer

This model is provided for research and educational purposes only. It is not intended for clinical decision-making. Always consult a qualified healthcare professional for medical diagnosis.

License

MIT. The underlying dataset has its own terms — see the dataset card before commercial use.

Author

Nasr Mohamed — AI Engineer 🤗 huggingface.co/nsr51324

nsr51324/Oral_Diseases_Image_Classification

🦷 Oral Diseases Image Classification