RootResolverAI

Use AI to do scoring of endodontic treatment outcomes Repository for the Python code used in the project

Background

Anonymized clips from periapical radiographs obtained and scored in the project Outcome of endodontic treatment at the Department of Endodontics, UiO[1] are used to train different machine learning models on PAI scoring. See repository EndodonticMeasurements [2] for details on the data aquisition.

Data

Image files

RGB images obtained by cropping radiographs to 224 x 224 pixels centered on the apex.

Scoring data

CSV-files with three columns: filename, PAI, weight

Models and training

Versions of the scripts here are snapshots. Several different amount of fine-tuning of layers, learnin rates and other hyperparameters were tested. Training data set is small and highly unbalanced:

EfficientNet(B7)[3]

Finetuning of last 3 blocks and classifier gave a validation accuracy of 0.56.

• Loss Function: Cross-entropy with per-sample weighting
• Optimizer: Adam (learning rate 0.0001, weight decay 1e-4)
• Scheduler: Cosine annealing for learning rate adjustment Program code for training and evaluation here.

ConvNext

More informatin be added

Program code for training and evaluation here.

ResNet-50

More informatin be added

Program code for training and evaluation here.

VGG19

Finetuning the entire pretrained model gave a validation accuracy of 0.46.

• Loss Function: Cross-entropy with per-sample weighting
• Optimizer: Adam (learning rate 0.0001, weight decay 1e-4)
• Scheduler: Cosine annealing for learning rate adjustment Program code for training and evaluation here.

References

1. Outcome of endodontic treatment at the Department of Endodontics, UiO
2. GitHub repository: EndodonticMeasurements
3. Tan, Mingxing, and Quoc V. Le. "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks." Proceedings of the 36th International Conference on Machine Learning (ICML), 2019