Fully automatic third molar staging from panoramic radiographs

In forensic dental age estimation practice, third molar development is the standard age predictor for sub-adults. The aim is to fully utilize the potential of deep learning to use panoramic radiographs (OPGs) directly for third molar staging, thereby bypassing time consuming manual interactions. Method. The data consists of 20 OPGs for each sex and for each developmental stage. A modified Demirjian staging technique was used by three forensic odontologists in consensus as a reference for the automatic stage allocation algorithm. In a first step of the automatic stage allocation, a 448x448 ROI around the geometrical center of the third molar is detected with a convolutional neural network (CNN). In a second step, a CNN is trained to segment the third molar within the ROI. In a third step, the segmentation is used to mask the original ROI by multiplication and the result is used as input for a CNN, which classifies the third molar into the corresponding developmental stage. This sequence was validated using five-fold cross-validation. Results. The geometrical center of the third molar was localized withan average Euclidean distance of 63.3 pixels. Within the detected ROIs, third molars were segmented with an average Dice score of 0.93. The stage classification from the automatically detected and masked ROI reached an accuracy of 0.54, a mean absolute error of 0.69 and a linear weighted kappa of 0.79. Conclusion. Despite the relatively limited number of subjects to learn from, this work showed promising results compared to manual stage allocation