Atlas-based neuroinformatics via MRI: harnessing information from past clinical cases and quantitative image analysis for patient care

Abstract

With the ever-increasing amount of anatomical information radiologists have to evaluate for routine diagnoses, computational support that facilitates more efficient education and clinical decision making is highly desired. Despite the rapid progress of image analysis technologies for magnetic resonance imaging of the human brain, these methods have not been widely adopted for clinical diagnoses. To bring computational support into the clinical arena, we need to understand the decision-making process employed by well-trained clinicians and develop tools to simulate that process. In this review, we discuss the potential of atlas-based clinical neuroinformatics, which consists of annotated databases of anatomical measurements grouped according to their morphometric phenotypes and coupled with the clinical informatics upon which their diagnostic groupings are based. As these are indexed via parametric representations, we can use image retrieval tools to search for phenotypes along with their clinical metadata. The review covers the current technology, preliminary data, and future directions of this field.

Figure 1

Relationship of the level of structural granularity and quantification strategies to define corresponding brain locations across different subjects. The lowest granularity level is the entire brain, in which one region of interest is defined across subjects. The highest granularity level is one voxel. In this case, for each arbitrary voxel chosen in one brain, a corresponding voxel is defined in the other brain. For voxel-based analysis, all the voxels of one brain are mapped to all the voxels of another brain. To perform a structure-level analysis, a cluster of voxels is defined, either manually or using an automated parcellation tool.

Figure 2

Different criteria define structures inside the brain. The same brain can be parcellated into various structures based on five different criteria: (a) classical structural units, (b) vascular territory, (c) anatomical connectivity, (d) functional connectivity, and (e) cytoarchitecture. (f) Age-dependent structural definition is also important because the contrasts and shapes of brain structures may vary with the age of the patient, and thus an adult atlas may not be applicable to pediatric or geriatric populations.

Figure 3

Whole-brain parcellation based on a preparcellated atlas. An atlas of choice (a), which defines the parcellation criteria, is elastically Warped to an individual image (e) and automatically parcels the entire brain into various structures. The images shown in panels b through d describe the transformation steps for automated parcellation. As the atlas is applied to many clinically normal and abnormal cases, the quantification results, such as the structural volumes of each parcel, are stored in the atlas, from which the average and the standard deviation of normal cases at each age can be characterized (f) (104).

Figure 4

Examples of image-matrix conversions. The multiple raw MR images (e.g., T2, FA, MD) are parcellated into approximately 200 structures, from which their size, T2, FA, and MD values are then quantified. This converts the three MR images into a 4 × 200 standardized matrix. If we have the same matrices from age-matched control subjects, the amount of deviation from the normal average (z-score) can be calculated. Examples include patient brains with focal and diffuse abnormalities and controls. These standardized and quantitative matrices are directly comparable across subjects and are searchable in the database. Abbreviations: FA, fractional anisotrophy; GM, gray matter; L, left; MD, mean diffusivity; R, right; SD, standard deviation; WM, white matter.

Figure 5

Principal component analysis (PCA) of anatomical matrices of primary progressive aphasia (PPA) patients and age-matched controls. The first three principal component (PC) axes account for 50.3% of observed variability in the anatomical matrices. Although the two groups are segregated in this space, there is a substantial amount of overlap. This could be due to a failure of the anatomical matrices to capture clinically important anatomical features. However, close inspection of outliers, such as the case indicated by a green arrow, reveals that their anatomy is, indeed, a typical of patients with the clinical phenotype of PPA. In this way, anatomical and clinical phenotypes of individuals can be systematically compared with population data.

Figure 6

A result of partial least squares discriminant analysis based on the results shown in Figure 4. In this analysis, the anatomical features in the anatomical matrices that maximize the separation of the two groups [patients (p) and controls (c)] are extracted from a training data set (n = 38) and tested for the ability to diagnose PPA in a test data set (n = 31). It is important to note that the purpose of this analysis was not to test the accuracy of automated diagnosis based on population data but rather to provide information about individual pathology status by incorporating both anatomical and clinical information. Specifically, if a new PPA patient is categorized in the p (diagnosis)/p (anatomy-based prediction) class, this patient has typical PPA-type anatomical features. If, however, a patient is diagnosed with PPA based on clinical symptoms but does not have PPA-like anatomical features (p/c class), the physician should be aware of the potential for misdiagnosis or for a special subtype of PPA with a different time course or outcome.