From Large to Small Organ Segmentation in CT Using Regional Context.

The segmentation of larger organs in CT is a well studied problem. For lungs and liver, state of the art methods reach Dice Scores above 0.9. However, these methods are not as reliable on smaller organs such as pancreas, thyroid, adrenal glands and gallbladder, even though a good segmentation of these organs is needed for example for radiotherapy planning. In this work, we present a new method for the segmentation of such small organs that does not require any deformable registration to be performed. We encode regional context in the form of anatomical context and shape features. These are used within an iterative procedure where, after an initial labelling of all organs using local context only, the segmentation of small organs is refined using regional context. Finally, the segmentations are regularised by shape voting. On the Visceral Challenge 2015 dataset, our method yields a substantially higher sensitivity and Dice score than other forest-based methods for all organs. By using only affine registrations, it is also computationally highly efficient.