Profile Scale-Spaces for Multiscale Image Match

Anatomical objects often have complex and varying image appearance at different portions of the boundary; and it is frequently a challenge even to select appropriate scales at which to sample the image. This motivates Bayesian image-match models which are both multiscale and statistical. We present a novel image-match model for use in Bayesian segmentation, a multiscale extension of image profile models akin to those in Active Shape Models. A spherical-harmonic based 3D shape representation provides a mapping of the object boundary to the sphere S, and a scale-space for profiles on the sphere defines a scalespace on the object. A key feature is that profiles are not blurred across the object boundary, but only along the boundary. This profile scalespace is sampled in a coarse-to-fine fashion to produce features for the statistical image-match model. A framework for model-building and segmentation has been built, and testing and validation are in progress with a dataset of 70 segmented images of the caudate nucleus. 1 Why are anatomical objects so hard to segment? Model-based segmentation has come a long way since Kass and Witkin’s original snakes [1], but segmentation of anatomical structures from real-world 3D medical images still presents some difficult challenges for automatic methods. Bayesian model-based segmentation balances a geometry prior, guided by a model of the expected object shape, against an image match likelihood, guided by a model of the expected image appearance around the object. Much has been done on the shape representation and prior; here we will focus on the image match model. In many objects, a simple gradient-magnitude based image-match model is insufficient. The profile of the image across the object boundary can vary significantly from one portion of the boundary to another. Some portions of the boundary might not even have a visible contrast, in which case the shape prior is needed to define the contour. In real-world medical images, the contrast-to-noise ratio is often low, and models need to be robust to image noise. In our study, one of the applications we focus on is the caudate nucleus in the human brain. From a partnership with our Psychiatry department, we have access to over 70 high-resolution MRIs (T1-weighted, 1x1x1mm) with high-quality manual expert segmentations of both left and right caudates. The manual raters, ? Supported by NIH-NCI P01 EB002779. having spent much effort on developing a reliable protocol for manual segmentation, indicate some of the challenges in caudate segmentation, which motivate a multiscale statistical image-match model for automatic methods. Portions of the boundary of the caudate can be localized with standard edge detection (provided the appropriate scales are chosen). However, there are also nearby false edges which may be misleading. In addition, where the caudate borders the nucleus accumbens, there is no contrast at the boundary; the manual raters use a combination of shape prior and external landmarks to define the boundary. Figure 1 shows the challenge. The caudate and nucleus accumbens are distinguishable on histological slides, but not on MRI of this resolution. Fig. 1. Coronal slice of the caudate: original T1-weighted MRI (left), and segmented (middle). Right and left caudate are shown shaded in green and red; left and right putamen are sketched in yellow, laterally exterior to the caudates. Nucleus accumbens is sketched in red outline. Note the lack of contrast at the boundary between the caudate and the nucleus accumbens, and the fine-scale cell bridges between the caudate and the putamen. At right is a 3D view of the caudate and putamen relative to the ventricles. Another “trouble-spot” for the caudate is where it borders the putamen; there are “fingers” of cell bridges which span the gap between the two. The scale of the object structure relative to the scale of the image noise may swamp single-scale image-match models with the noise. Many other segmentation tasks in medical images present challenges similar to the caudate; in automatic segmentation methods, this motivates image-match models which are statistically trained and multiscale. This paper focuses on the image match likelihood model, not on the shape prior. The shape prior we use is via the statistical spherical harmonics shape model [2].