AlphaDent
Dataset links
- Dataset on zenodo, kaggle, huggingface, github
Train
- Download dataset and put in the folder with this code. Then fix path in
yolo_seg_train.yamlif needed. - Then you can train with following script:
python3 train.py --dataset_config ./AlphaDent/yolo_seg_train.yaml --batch_size 16 --epochs 100 --image_size 640
Results of training will be stored in folder ./yolo_seg_x_proj_640.
Pretrained weights
There are 3 different pretrained wights available:
- Yolo_v8x, 9 classes and 640 input. Download: Link
- Yolo_v8x, 9 classes and 960 input. Download: Link
- Yolo_v8x, 4 classes and 960 input. Download: Link
Validation
Validation will run model with validation data and output metrics.
python3 valid.py --weights './weights/yolov8x_AlphaDent_9_classes_640px.pt' --dataset_config './AlphaDent/yolo_seg_train.yaml' --batch_size 16 --epochs 100 --image_size 640
Inference
If you have new dental photos for which you want to obtain predictions you can use inference script.
python3 inference.py --weights './weights/yolov8x_AlphaDent_9_classes_640px.pt' --input_path './AlphaDent/images/test/' --output_path './output/' --batch_size 16 --image_size 640
Useful scripts
Convert 9 classes dataset to 4 classes
python3 utils/convert_9_classes_dataset_to_4_classes.py --input_path './AlphaDent/' --output_path './AlphaDent_4_classes/'
Draw Yolo annotations
python3 utils/draw_annotations.py --input_path './AlphaDent/' --output_path './Draw_Annotations/'
Citations
If you find this work useful please cite:
@misc{sosnin2025alphadent,
title={AlphaDent: A dataset for automated tooth pathology detection},
author={Evgeniy I. Sosnin and Yuriy L. Vasilev and Roman A. Solovyev and Aleksandr L. Stempkovskiy and Dmitry V. Telpukhov and Artem A. Vasilev and Aleksandr A. Amerikanov and Aleksandr Y. Romanov},
journal={Computer Optics},
volume={49},
pages={6},
year={2025},
}