This document summarizes a study that developed a random forest classifier to improve bone tumor segmentation on bone scans. The classifier uses contextual information like offset vectors from tumor landmarks and symmetry features, in addition to intensity and texture features. It was evaluated on bone scans from prostate cancer patients and showed improved segmentation performance over an existing rule-based method, with the contextual features providing most of the discriminative power. The random forest classifier with contextual features increased the Jaccard index by 0.09 compared to a classifier without contexts, and by 0.08 compared to the previous rule-based method, better reducing false positives.