IADR Abstract Archives
Ensemble Learning for detection of periapical pathology and its relation to teeth on Orthopantomograms – A validation study
Objectives: 1. To develop a Deep Learning (DL) model for the segmentation of periapical pathology and its relation to teeth on Orthopantomograms (OPGs).
2. To validate its performance relative to the ground-truth laid down by human annotators.
3. To create an ensemble of our previously trained teeth segmentation and numbering model with the current lesion detection model.
Methods: 250 OPGs were manually annotated by subject experts to lay down the ground-truth for training a DL Artificial Intelligence (AI) model for segmentation of periapical pathology (shown in figure 1). An untrained U-net algorithm was trained and validated on the labelled dataset. Our previously trained model on teeth segmentation and numbering tasks was also further trained on 250 labelled OPGs to improve performance and accuracy of teeth segmentation and numbering tasks. Both models were then integrated via code composed specifically for the task of combining our previous and current algorithms running in parallel. This combination allowed for relation of periapical pathology to the causative tooth on OPG, creating an ensemble of all our AI algorithms.
Results: The performance of the existing teeth segmentation and numbering model was further improved as indicated by the following performance metrics including accuracy=98.1%, precision=91.8%, re-call=93.3%, F-1 score=92.5%, dice index=90.1% and Intersection over Union (IoU)=82.1% (visual results are shown in figure 2). The performance metrics of lesion segmentation carried out by the current model are as follows: accuracy=98.1%, precision=84.5%, re-call=80.3%, F-1 score=82.2%, dice index=75.2% and IoU=67.6%. The integrated visual results of both models combined are shown in figure 3.
Conclusions: Our ensemble performs the task of lesion detection and its relation to the causative tooth on OPG with comparable results to that laid down in the ground-truth as indicated by performance metrics.
2. To validate its performance relative to the ground-truth laid down by human annotators.
3. To create an ensemble of our previously trained teeth segmentation and numbering model with the current lesion detection model.
Methods: 250 OPGs were manually annotated by subject experts to lay down the ground-truth for training a DL Artificial Intelligence (AI) model for segmentation of periapical pathology (shown in figure 1). An untrained U-net algorithm was trained and validated on the labelled dataset. Our previously trained model on teeth segmentation and numbering tasks was also further trained on 250 labelled OPGs to improve performance and accuracy of teeth segmentation and numbering tasks. Both models were then integrated via code composed specifically for the task of combining our previous and current algorithms running in parallel. This combination allowed for relation of periapical pathology to the causative tooth on OPG, creating an ensemble of all our AI algorithms.
Results: The performance of the existing teeth segmentation and numbering model was further improved as indicated by the following performance metrics including accuracy=98.1%, precision=91.8%, re-call=93.3%, F-1 score=92.5%, dice index=90.1% and Intersection over Union (IoU)=82.1% (visual results are shown in figure 2). The performance metrics of lesion segmentation carried out by the current model are as follows: accuracy=98.1%, precision=84.5%, re-call=80.3%, F-1 score=82.2%, dice index=75.2% and IoU=67.6%. The integrated visual results of both models combined are shown in figure 3.
Conclusions: Our ensemble performs the task of lesion detection and its relation to the causative tooth on OPG with comparable results to that laid down in the ground-truth as indicated by performance metrics.
3874133_File000005.jpg
3874133_File000006.jpg
3874133_File000007.jpg