Probabilistic Brain Atlas Encoding Using Bayesian Inference

Abstract

This paper addresses the problem of creating probabilistic brain atlases from manually labeled training data. We propose a general mesh-based atlas representation, and compare different atlas models by evaluating their posterior probabilities and the posterior probabilities of their parameters. Using such a Baysian framework, we show that the widely used "average" brain atlases constitute relatively poor priors, partly because they tend to overfit the training data, and partly because they do not allow to align corresponding anatomical features across datasets. We also demonstrate that much more powerful representations can be built using content-adaptive meshes that incorporate non-rigid deformation field models. We believe extracting optimal prior probability distributions from training data is crucial in light of the central role priors play in many automated brain MRI analysis techniques.

Model-Based Segmentation of Hippocampal Subfields in Ultra-High Resolution In Vivo MRI

“…The model incorporates a prior distribution that makes predictions about where neuroanatomical labels typically occur throughout the image, and is based on the generalization of probabilistic atlases [2,3,4,5,11] developed in [12]. The model also includes a likelihood distribution that predicts how a label image, where each voxel is assigned a unique neuroanatomical label, translates into an MRI image, where each voxel has an intensity.…”

“…It has previously been demonstrated that the mesh's connectivity, reference position x r , and label probabilities α can be learned from a set of manually labeled example images [12]. The learning involves selecting the model that maximizes the probability of observing the example label images, or, equivalently, that minimizes the number of bits needed to encode them.…”

Prior Knowledge Driven Multiscale Segmentation of Brain MRI

“…Similarly to other MRI brain segmentation methods [1,2,3,4] we employ probabilistic anatomical atlases to determine the prior information. The atlas is constructed based on affine co-registrations of the manually labeled training set to the test subject, implemented using the publicly available AIR5.0 [15] registration algorithm with 12-parameter affine transformation.…”

“…The challenge in brain MRI segmentation is due to issues such as noise, intensity non-uniformity (INU), partial volume effect, shape complexity and natural tissue intensity variations. Under such conditions, incorporation of a priori medical knowledge, commonly represented in anatomical brain atlases by state-of-the-art studies [1,2,3,4] is essential for robust and accurate automatic segmentation.…”

Effects of Registration Regularization and Atlas Sharpness on Segmentation Accuracy

“…Twining et al [2] and Van Leemput [7] propose frameworks to find the least complex models that explain the image intensity and segmentation labels respectively in the training images. This is useful if the goal is to analyze the training images.…”