Magnetic resonance imaging segmentation of hippocampus using atlas selection by meta-information
Hippocampus segmentation in magnetic resonance imaging is an important procedure in many clinical situations, such as monitoring changes in patients with Alzheimer's disease. However, the manual delineation of this structure, in three-dimensional images, is a laborious task and prone to subjective interpretation of the health professional.
Some automated methods have been proposed in recent years. Much of these methods use pre-segmented templates, also known as atlas, which are aligned to the input image in the segmentation process. However, using a single standard atlas increases the difficulty targeting individuals that have non-normal anatomy, such as the elders and patients with AD. To achieve a good precision in these cases, without any manual intervention of the user, new methods employ techniques in which several different atlases are used.
The alignment of these atlases with the image, leads to a high computational cost. This work proposes employing an atlas selection technique by meta-information in order to choose the ideal template for an individual, enabling low computational cost segmentation technique. The results obtained, by testing 350 individuals, in various clinical conditions and ages, showed that the use of atlas selection significantly increases segmentation accuracy, when compared to a method using a default atlas, while keeping the computational cost low. The relevance of three selection parameters - medical condition, age and gender - has been evaluated and confirmed by the test suite.