Generative modeling of medical images
Yael Balbastre
University College London
Add to Calendar
2019-06-26 16:00:00
2019-06-26 17:00:00
America/New_York
Generative modeling of medical images
Probabilistic inference consists of estimating a probabilitydistribution based on a limited number of randomly sampledobservations. When these observations are images, Euclidean inference(assuming no prior covariance among voxels) often fails to estimate arepresentative distribution of the data. This problem can be overcomeby accounting for two characteristics of images: first, theirintrinsic smoothness, which is captured by a local covariance amongvoxels; and second, their topology, which captures the fact that theobjects represented in the images are invariant under some families oftransformations (e.g., multiplicative or additive changes ofappearance, affine or non-linear spatial deformations).In this talk I will show that a set of images can be described by amean and a distribution of transformations (of a given type), suchthat a single transformation from the distribution would map the meanimage to a sample from the set of images, and that the particulartransformation type depends on the nature of the variability to bemodeled. I will show two practical applications capitalizing on thisframework: the estimation of sensitivity fields in multi-coil MRacquisitions, and the estimation of brain templates in computationalanatomy. I will then show that by extending the model of priorcovariance from capturing local smoothness only, to having anon-stationary form, more structured deviations from the mean imagecan be captured. This concept will be applied to the estimation ofshape and appearance variability in the human brain.
32-D451
Belfer
sarah_donahue@hks.harvard.edu