| Summary: | Deformable templates, or atlases, are images, often labelled, that represent a typical anatomy for a population. They are commonly used in medical image analysis for population studies and computational anatomy tasks. Practitioners use image alignment techniques to compare the subject scan and the template. Unfortunately, developing a template is a computationally expensive process with existing methods. Usually, at most one template is available per population of images or anatomy. As a results, analysis is often conducted with sub-optimal templates. In this thesis, we propose a machine learning framework that uses convolutional alignment neural networks to efficiently create both unconditional and conditional templates and the corresponding label maps. We demonstrate our method on a large 3D brain MRI dataset. This is particularly relevant in medical image analysis where templates are difficult to build. We show that this framework can learn sharp templates representative of the population. These templates are representative of the population. Moreover, they can leverage label maps when available. Our method enables rapid registration of any brain image to our template. Moreover, our method has the options of producing representative conditional templates, given subject specific attributes.
|
|---|