PAI-meets-AI: Open-Weight Ensemble for Periapical Index Scoring

This repository contains the official pre-trained model weights for the study "Development and Internal Validation of CNN-Based Models for Five-Class PAI Scoring".

We provide three distinct, fine-tuned Deep Learning architectures—ResNet50, EfficientNet-B3, and ConvNeXt-Tiny—which can be used individually or combined into an Equal-Weight Soft-Voting Ensemble.

The Periapical Index (PAI) is a standardized 1–5 scale used for assessing apical periodontitis on dental radiographs. These models were trained on a dataset of 17,549 expert-annotated intraoral radiographs to automate this assessment.

• 💻 Source Code & GUI: https://github.com/geraldOslo/PAI-meets-AI
• 📄 Paper: [Link to paper/preprint if available]

Available Models

We release the weights for the best-performing checkpoint of each architecture.

File Name	Architecture	Input Size	Description
efficientnet-b3_best.pth	efficientnet_b3	300x300	Best single model in our study (QWK 0.769).
resnet50_best.pth	resnet50	224x224	High sensitivity for healthy cases.
convnext-tiny_best.pth	convnext_tiny	224x224	Modern Transformer-style CNN.

Note: An older legacy checkpoint (pai-meets-ai_...) is preserved for reproducibility but the specific named files above are recommended.

Normalization Details

All models were trained using standard ImageNet normalization statistics:

• Mean: [0.485, 0.456, 0.406]
• Std: [0.229, 0.224, 0.225]

🚀 How to Use

The checkpoints contain the model_state_dict along with the necessary configuration (model_name, num_classes) to reconstruct the architecture using timm.

Intended Use & Limitations

Research Only: These models are intended for research and educational purposes. They are not certified medical devices (SaMD) and should not be used for primary clinical diagnosis.

Input Data: The models are trained on cropped periapical radiographs centered on the tooth apex. Feeding full uncropped radiographs will lead to incorrect results.

License: The weights are released under the MIT License, allowing for open use, modification, and distribution with attribution.

Data Privacy

The original dataset containing 17,549 clinical radiographs is private and protected by GDPR/health data regulations. It cannot be shared. However, these released weights act as a privacy-preserving "condensed representation" of that dataset, enabling others to benefit from the learned features without accessing patient data. Citation

If you use these weights in your research, please cite our paper: code Bibtex

@software{Torgersen_PAI-meets-AI_2025, author = {Torgersen, Gerald and [Co-Authors]}, title = {{PAI-meets-AI: Open-Weight Deep Learning Models for PAI Scoring}}, year = {2025}, publisher = {GitHub/HuggingFace}, url = {https://github.com/geraldOslo/PAI-meets-AI} }