Evaluation of Cross-Protocol Stability of a Fully Automated Brain Multi-Atlas Parcellation Tool

Zifei Liang, Xiaohai He, Can Ceritoglu, Xiaoying Tang, Yue Li, Kwame S. Kutten, Kenichi Oishi, Michael I. Miller, Susumu Mori, Andreia V. Faria, Huiguang He
2015 PLoS ONE  
Brain parcellation tools based on multiple-atlas algorithms have recently emerged as a promising method with which to accurately define brain structures. When dealing with data from various sources, it is crucial that these tools are robust for many different imaging protocols. In this study, we tested the robustness of a multiple-atlas, likelihood fusion algorithm using Alzheimer's Disease Neuroimaging Initiative (ADNI) data with six different protocols, comprising three manufacturers and two
more » ... agnetic field strengths. The entire brain was parceled into five different levels of granularity. In each level, which defines a set of brain structures, ranging from eight to 286 regions, we evaluated the variability of brain volumes related to the protocol, age, and diagnosis (healthy or Alzheimer's disease). Our results indicated that, with proper pre-processing steps, the impact of different protocols is minor compared to biological effects, such as age and pathology. A precise knowledge of the sources of data variation enables sufficient statistical power and ensures the reliability of an anatomical analysis when using this automated brain parcellation tool on datasets from various imaging protocols, such as clinical databases. (SM). of human judgment is considered more reliable than automated detection tools. Even after tumor and/or stroke are ruled out, MRI data still contain a wealth of anatomical information that could be medically informative. These include changes in volume or shape, based on which the level of atrophy of specific brain structures could be evaluated, and changes in intensity, such as hyper intense ischemic lesions seen in T2-weighted and Fluid Attenuated Inversion Recovery (FLAIR) images. Unlike tumor and stroke, in which abnormal features appear that should not exist in normal brains, the shape and intensity changes caused by many neurodegenerative diseases, especially in the early stage or preclinical stage, are an extension of a continuum from a normal range of anatomical variability and age-dependent changes. With the sheer number of brain structures and image voxels, it could be argued that the human ability to capture the anatomical features and relate these to clinical outcomes is limited. Indeed, our current knowledge about the relationship between anatomical features and clinically important information, such as diagnosis, functional loss, or prognosis, is not strong enough to allow MRI to play more than a small role in medical decision-making and patient care for neurodegenerative diseases and dementia populations [1] [2] [3] . A widely used research paradigm in brain MRI is to use quantitative image analysis to quantify anatomical features, perform correlation analysis with clinical information, and eventually find important features (i.e., biomarkers or surrogate markers) that cannot be well appreciated by human perception alone. This paradigm has supported numerous studies in the past, in which an analysis based on image normalization, such as voxel-based analysis, was used in many studies (for review, see [4] ). Based on specific anatomical features that are linked to a patient group of interest, successful discrimination of diseases, such as Alzheimer's disease, has been reported [5], including studies using ADNI data [6] [7] [8] . While successful, the application of these approaches in routine clinical practice, however, would face several challenges. This is partly due to fundamental differences in the study design. Specifically, in research settings, the patient population is highly homogenized (so that the abnormality exists in common anatomical locations among the patient group), image protocols are consistent (so that subtle changes can be consistently quantified), a control population of high quality is available (so that the normal range of anatomical variability and bias in the population averages are minimized), and, eventually, group comparisons can be performed with proper statistical power. All these factors are not the case in clinical practice. Indeed, the initial stratification of the highly heterogeneous patient population is one of the most important missions of clinical MRI, and eventually, a judgment must be made for each individual patient. The heart of this clinical paradigm is the knowledge-versus-individual, not the group-versus-group study design, in which knowledge information from all individuals is retained without a reduction into group-based statistics (group-aggregated statistics cannot be used for a heterogeneous population) [9] . If we want to create a quantitative knowledge database of anatomical phenotypes, there are several interesting questions to be addressed. First, it is uncertain whether raw images with more than one million voxels are suitable as the contents of such a database; the sheer amount of the noisy voxel information could severely hamper our ability to store, search, and analyze the anatomical contents and perform a group vs. individual analysis. Second, such a large database inevitably needs to contain data from multiple sources and the tools that extract anatomical information need to be robust for a reasonable range of variability in scanning protocols. As long as the protocol effects cannot be completely eliminated, it is also important to know the range of measurement variability that would enable estimation of the abnormality-detection power, which would include both biological and artifactual (protocol differences, hardware performance, quantification accuracy) contributions. In this paper, we have extended past efforts to convert the dense imagery information into a set of structural representations by brain parcellation tools [10] [11] [12] [13] [14] . We are especially Cross-Protocol Stability of a Brain Multi-Atlas Parcellation Tool PLOS ONE |
doi:10.1371/journal.pone.0133533 pmid:26208327 pmcid:PMC4514626 fatcat:3esmuc5pt5dc3catxheo6sg22q