Search

Spatial data augmentation is a standard technique for regularizing deep segmentation networks that are tasked with localizing medical abnormalities. However, a typical spatial augmentation scheme is built upon ad hoc selections of spatial transformation parameters which are not determined by the data set and therefore may not capture spatial variations in the data. For segmentation networks trained in the low-data regime, these ad hoc transformation techniques often fail to encourage better generalization. To address this problem, we propose a variational autoencoder framework for spatial data augmentation. We show how this framework provides a natural, data-driven approach to probabilistic, instance-specific spatial augmentation. Further, we observe that U-Nets trained on data augmented using this framework compare favorably with U-Nets trained using standard spatial augmentation methods.

Domain shift is a common problem in machine learning and medical imaging. Currently one of the most popular domain adaptation approaches is the domain-invariant mapping method using generative adversarial networks (GANs). These methods deploy some variation of a GAN to learn target domain distributions which work on pixel level. However, they often produce too complicated or unnecessary transformations. This paper is based on the hypothesis that most domain shifts in medical images are variations of global intensity changes which can be captured by transforming histograms along with individual pixel intensities. We propose a histogram-based GAN methodology for domain adaptation that outperforms standard pixel-based GAN methods in classifying chest x-rays from various heterogeneous target domains.

In the study of shapes of human organs using computational anatomy, variations are found to arise from inter-subject anatomical differences, disease-specific effects, and measurement noise. This paper introduces a stochastic model for incorporating random variations into the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework. By accounting for randomness in a particular setup which is crafted to fit the geometrical properties of LDDMM, we formulate the template estimation problem for landmarks with noise and give two methods for efficiently estimating the parameters of the noise fields from a prescribed data set. One method directly approximates the time evolution of the variance of each landmark by a finite set of differential equations, and the other is based on an Expectation-Maximisation algorithm. In the second method, the evaluation of the data likelihood is achieved without registering the landmarks, by applying bridge sampling using a stochastically perturbed version of the large deformation gradient flow algorithm. The method and the estimation algorithms are experimentally validated on synthetic examples and shape data of human corpora callosa.