Automatic Segmentation of Tooth Images: Optimization of Multi-parameter Image Processing Workflow

Abstract

The development of specific algorithms in image processing are usually related to dataset characteristics. Those characteristics will influence the number of instructions required to solve a problem. Normally, the more complex a set of instructions is, the more parameters need to be set. Dealing with such degrees of freedom, sometimes leading to subjective decision making, is time-consuming and frequently leads to errors or sub-optimal results of the developed model. Here, we deal with a model for segmentation of masks of tooth images containing a pattern of bands called Hunter-Schreger Bands (HSB). They appear on tooth surface when lit from the side. This segmentation process is only one step of a pipeline whose overall goal is human biometric identification to be used, e.g., in forensics. The segmentation algorithm, which exploits the anisotropy of the image, uses several parameters and choosing the optimal combination of them is challenging. The aim of this work was to utilize visual data analysis tools to optimize the chosen parameters and to understand their influence on the performance of the algorithm. Our results reveal that a slightly better combination of parameter values can be found starting from the experimentally determined initial parameters. This approach can be repeatedly performed to achieve even better parameterizations. To more deeply understand the influence of the parameters on the final result, more sophisticated visual interaction tools will be explored in future work.