Rhesus Teeth Segmentation by Use of Multi-Atlas Library


A fully automated tooth segmentation tool was developed to aid in the detection and segmentation of rhesus monkey teeth in CT images. The tool utilized multi-atlas segmentation at several resolutions and levels of zoom to minimize the processing time while still maintaining high accuracy in the segmentations.

A example input data set showing both the full field-of-view low resolution and the focused tooth high resolution image and the four regions to segment.

A example input data set showing both the full field-of-view low resolution and the focused tooth high resolution image and the four regions to segment.

Twelve CT images were provided with regions of interest manually defined to act as a reference library. The four teeth segmented were the left incisor, the left canine, and two molars on the left side. The multi-atlas segmentation process goes as follows:

1) Register each reference data set to the input image using an affine transform followed by a deformable transform. Apply the same transforms to the region of interest data.
2) Average a number of the transformed reference ROIs together to construct a probability map in on the input image of each region of interest.
3) Threshold the probability map to build a deterministic output region of interest.

Multi-Atlas segmentation was run on the entire field of view at a low resolution for all 4 regions. A bounding box for each output region was determined and the region was extracted at high resolution for each region. Multi-Atlas segmentation was then run again on each focused high resolution region to determine the final segmentation.


Example workflow of the segmentation for the incisor of an input data set

Example workflow of the segmentation for the incisor of an input data set


Overview of the 4 computed metrics for evaluation of the tool

Overview of the 4 computed metrics for evaluation of the tool


Performance was evaluated by running each member of the reference library through the tool while excluding the input from the reference library. The output automated segmentations were compared to the provided manual segmentations at a voxel-by-voxel level and metrics such as the Dice coefficient, Jaccard index, sensitivity and precision were computed

Example results averaged over all 12 input data sets for the Incisor

Example results averaged over all 12 input data sets for the Incisor


The tool was packaged in a way to run in batch from the command line as a shell script. It utilized several custom-built programs as well as VivoQuant® to generate the output results.