Automatic organ localization

A computer-implemented method of training a regression tree to localize an organ in a medical image, the method comprising:

• receiving at least one training medical image comprising a plurality of image elements and a predefined organ location;
• selecting a node of the regression tree and providing the plurality of image elements to the node;
• generating parameters defining a test performed on each image element at the node that determines which subsequent node of the tree each image element is passed to;
• repeating the step of generating parameters for each subsequent node until each image element reaches one of a plurality of leaf nodes of the regression tree;
• at each leaf node, generating a probability distribution based on a distance from each image element reaching that leaf node to the organ location;
• and
• storing the probability distribution associated with each leaf node at a storage device.

A device for automatically localizing an organ in a three-dimensional medical volumetric image, comprising:

• a communication interface arranged to receive the three-dimensional medical volumetric image;
• a storage device arranged to store regression forest comprising a plurality of distinct trained regression trees;
• and
• a processor arranged to:
• apply each voxel of the three-dimensional medical volumetric image to each trained regression tree to compute a plurality of probability distributions for each image element relating to a distance from that voxel to each wall of a three-dimensional bounding box surrounding the organ;
• select a subset of the probability distributions;
• and aggregate these selected probability distributions to compute a localization estimate for the bounding box surrounding the organ.