Genetic programming for landmark detection in cephalometric radiology images

This paper describes the use of genetic programming to evolve object detection programs for craniofacial features in digital X-rays. The evolved programs use a feature set of pixel statistics of regions customised to the shapes of landmark idiosyncrasies. The features are obtained from a square input window centred on the landmark and large enough to contain key landmark features. The detection program is applied, in moving window fashion, across the X-ray and the outpu7t of the program is interpreted as the presence/absence of the landmark at each position. During training a weighted combination of the detection rate and the false alarm rate is used as the fitness function. The method was tested in 4 landmark points, ranging from relatively easy to very difficult. Detection performance on the easier points was excellent and the performance on the very difficult point was quite good and the results suggest that a more careful crafting of the regions shapes for the difficult point will lead to better detection. We believe that the methodology can be used successfully on other difficult real world detection problems.