Prof John Ashburner University College London, UK |
Deformation Models |
All scientific questions are framed within some form of model, and often involve identifying the model or family of models that encodes the better explanation of the data. Deformations form an important component of models of biological images. The pervasive view is that image registration ought to be a tool that is run as part of a processing pipeline, such that the output from one step becomes the input for the next, and so on. In the first part of my talk, I will try to argue against this opinion, and suggest that the biological sciences may be better served by formulating generative models of primary data. In the second part, I will show some of my experiments with diffeomorphic deformations. When constructing models, physicists base many of their assumptions on invariances. While this leads to more parsimonious explanations of data and therefore more accurate predictions, it often leads to greater computational complexity. Similarly, diffeomorphic deformation models may be a bit more computationally expensive to fit, but they can provide a much simpler encoding of the image data.