Automatic localization of locally similar structures based on the scale-widening random regression forest

Selection of set of training pixels and feature range show to be critical scale-related parameters with high impact on results in localization methods based on random regression forests (RRF). Trained on pixels randomly selected from images with long range features, RRF captures the variation in landmark location but often without reaching satisfying accuracy. Conversely, training an RRF with short range features in a landmark's close surroundings enables accurate localization, but at the cost of ambiguous localization results in the presence of locally similar structures. We present a scale-widening RRF method that effectively handles such ambiguities. On a challenging hand radiography image data set, we achieve median and 90th percentile localization errors of 0.81 and 2.64mm, respectively, outperforming related state-of-the-art methods.