IADR Abstract Archives
Artificial Intelligence as a Decision-Making Tool in Forensic Dentistry: a Pilot Study
Objectives: Expert determination of the third molar maturity index (I3M) constitutes one the most common approach for dental age estimation. This work aimed to replicate this index by combining artificial intelligence solutions with topological image analysis to help and support expert decision-making.
Methods: The dataset consisted of 456 images from France and Ouganda. Two deep-learning supervised approaches (Mask RCNN, U-Net) were compared on mandibular third molar radiographs, leading to a two-part instance segmentation (apical and coronal). Then, in order to determine the tooth height and apex openness, two topological data analysis approaches were compared on the inferred mask: one with a deep learning component (TDA-DL), one without (TDA).
Results: Regarding mask inference, U-Net had a better accuracy (mean intersection over union metric (mIoU)) 91.2% compared to Mask R-CNN 83.8%. The combination of U-Net with TDA or TDA-DL to compute the I3M score revealed satisfying results in comparison with a dental forensic expert. The mean ± SD absolute error was 0.04 ± 0.03 for TDA, and 0.06 ± 0.04 for TDA-DL. The Pearson correlation coefficient of the I3M scores between the expert and a U-Net model was 0.93 when combined with TDA and 0.89 with TDA-DL.
Conclusions: This fully-automated solution combining a deep learning and a topological approach led to a 95% accuracy in comparison with an expert. Consequently, this proof-of-concept illustrates the feasibility of a fully-automated I3M computation that could be used to assist dental expert or to enhance existing age estimation algorithms.
Methods: The dataset consisted of 456 images from France and Ouganda. Two deep-learning supervised approaches (Mask RCNN, U-Net) were compared on mandibular third molar radiographs, leading to a two-part instance segmentation (apical and coronal). Then, in order to determine the tooth height and apex openness, two topological data analysis approaches were compared on the inferred mask: one with a deep learning component (TDA-DL), one without (TDA).
Results: Regarding mask inference, U-Net had a better accuracy (mean intersection over union metric (mIoU)) 91.2% compared to Mask R-CNN 83.8%. The combination of U-Net with TDA or TDA-DL to compute the I3M score revealed satisfying results in comparison with a dental forensic expert. The mean ± SD absolute error was 0.04 ± 0.03 for TDA, and 0.06 ± 0.04 for TDA-DL. The Pearson correlation coefficient of the I3M scores between the expert and a U-Net model was 0.93 when combined with TDA and 0.89 with TDA-DL.
Conclusions: This fully-automated solution combining a deep learning and a topological approach led to a 95% accuracy in comparison with an expert. Consequently, this proof-of-concept illustrates the feasibility of a fully-automated I3M computation that could be used to assist dental expert or to enhance existing age estimation algorithms.