| Summary: | The increasing advancements in the machine learning field combined with the performance disparity of manual data processing has led to the application of automation to clinical tasks. This study aimed to explore the detection performance of various object detection machine learning algorithms for dental attributes like the root canal and implant. A total of 700 dental panoramic images were utilised for this study. LabelMe, an image annotation tool, was chosen to complete the annotation of the images due to its ease of use and stability as compared to other image annotation tools. The machine learning object detection methods of interest were the TensorFlow 2 based methods of Faster R-CNN and SSD due to their current reputation of being widely preferred for object detection use cases. The Faster R-CNN model employs the ResNet 101 backbone network as it is the current state-of-the-art object detection model with performance better than other alternatives. The SSD model employs the MobileNet V2 FPNLite backbone network for its simplicity, which could allow analysis to be done on mobile devices. The 34 classes, defined as 32 teeth based on the FDI (Federation Dentaire Internationale) numbering system, root canal and implant, were detectable by the SSD model. The mAP achieved was 0.0113 while the Total Loss was 2.333. In comparison with the Faster R CNN model previously studied, which had a mAP of 0.4391 and a Total Loss of 1.552, the SSD model was found to have quicker detection but compromised accuracy. The detection scores of the model range from 45% to 90% on the test images which signifies that the detection performance still requires improvements and are not comparable to that of a dental practitioner. This concludes that the Faster R-CNN model is the more accurate dental attributes detection solution.
|
|---|