Post-processing of Anatomical Landmark Detection: Distance Error Reduction by Pictorial Structure Matching-Based Method

Abstract

The detection of anatomical landmarks (LMs) often plays a key role in medical image analysis. We have been studying an automatic detection method for multiple LMs from human torso CT data. In our latest experiments on the detection of 181 LMs from 39 human torso CT data, the sensitivity was 97.4% and the average distance error of the detected LM locations was 8.01 mm. Although about 80% of the LM detection results had a distance error of less than 10 mm, there is still room for improvement in the detection. In this study, we introduce a post processing method to refine LM locations, which are detected by our previous method. The proposed refinement method based on pictorial structure matching is carried out using a pictorial structure model including the following information: the local appearance of the refinement target LM, the spatial distribution of the target LM and support LMs including the spatial association with the target LM, and the local appearance of the support LMs. The location with the maximum likelihood calculated by the model is defined as the refined LM location. The proposed method was evaluated with 190 detected locations of 5 LMs in 39 human torso CT data. By applying the proposed refinement, the distance errors were reduced in 137 LM locations, which is 72.1 % of the total. The average distance error, which was originally 15.3 mm, was reduced to 9.7 mm. These results showed the potential of the proposed method for reducing the distance errors of detected LM locations.