Improving point correspondence in cephalograms by using a two-stage rectified point transform.

BACKGROUND

An improved point correspondence method was developed for automatically detecting two-dimensional cephalometric landmarks. The proposed method uses a two-stage rectified point transform: the global correspondence of interest points between two images and the local correspondence of landmarks.

METHOD

In the first stage, point-to-point matching pairs were established using local corner point features. The matched points on an input image were treated as a set of transformations, with varying directions and magnitudes, from the template image. Similarity of the transformation vectors was achieved through rectification to exclude vectors that deviated widely from the statistical mean. Rectification attempted to remove noise and irrelevant matched points. In the second stage, the point correspondences were fine-tuned within the regional centers of the landmarks, which were classified into three categories-corners, edges, and structural points-and each category was fine-tuned using a different strategy. Correspondence was performed by evaluating the shortest Euclidean distance between the point descriptors of the template and test images.

RESULTS

The correspondence results of 20 orthodontic landmarks were compared with those identified by dental professionals on 80 digital cephalograms collected from a dental clinic. The proposed method detected both hard and soft tissue landmarks with mean error distances of 1.63mm, compared with the 2-mm standard reported by previous studies.

CONCLUSIONS

This study enhanced the point correspondence technique for cephalometric landmarking. Using the proposed method, users can preferentially and flexibly add and remove landmarks on a template before correspondence without intensive image pretraining.