D.L. Collins

Automatic 3D intersubject registration of MR volumetric data in standardized Talairach space.

Objective In both diagnostic and research applications, the interpretation of MR images of the human brain is facilitated when different data sets can be compared by visual inspection of equivalent anatomical planes. Quantitative analysis with predefined atlas templates often requires the initial alignment of atlas and image planes. Unfortunately, the axial planes acquired during separate scanning sessions are often different in their relative position and orientation, and these slices are not coplanar with those in the atlas. We have developed a completely automatic method to register a given volumetric data set with Talairach stereotaxic coordinate system. Materials and Methods The registration method is based on multiscale, three-dimensional (3D) cross-correlation with an average (n > 300) MR brain image volume aligned with the Talairach stereotaxic space. Once the data set is resampled by the transformation recovered by the algorithm, atlas slices can be directly superimposed on the corresponding slices of the resampled volume. The use of such a standardized space also allows the direct comparison, voxel to voxel, of two or more data sets brought into stereotaxic space. Results With use of a two-tailed Student t test for paired samples, there was no significant difference in the transformation parameters recovered by the automatic algorithm when compared with two manual landmark-based methods (p > 0.1 for all parameters except y-scale, where p > 0.05). Using root-mean-square difference between normalized voxel intensities as an unbiased measure of registration, we show that when estimated and averaged over 60 volumetric MR images in standard space, this measure was 30% lower for the automatic technique than the manual method, indicating better registrations. Likewise, the automatic method showed a 57% reduction in standard deviation, implying a more stable technique. The algorithm is able to recover the transformation even when data are missing from the top or bottom of the volume. Conclusion We present a fully automatic registration method to map volumetric data into stereotaxic space that yields results comparable with those of manually based techniques. The method requires no manual identification of points or contours and therefore does not suffer the drawbacks involved in user intervention such as reproducibility and interobserver variability.

Design and construction of a realistic digital brain phantom

After conception and implementation of any new medical image processing algorithm, validation is an important step to ensure that the procedure fulfils all requirements set forth at the initial design stage. Although the algorithm must be evaluated on real data, a comprehensive validation requires the additional use of simulated data since it is impossible to establish ground truth with in vivo data. Experiments with simulated data permit controlled evaluation over a wide range of conditions (e.g., different levels of noise, contrast, intensity artefacts, or geometric distortion). Such considerations have become increasingly important with the rapid growth of neuroimaging, i.e., computational analysis of brain structure and function using brain scanning methods such as positron emission tomography and magnetic resonance imaging. Since simple objects such as ellipsoids or parallelepipedes do not reflect the complexity of natural brain anatomy, the authors present the design and creation of a realistic, high-resolution, digital, volumetric phantom of the human brain. This three-dimensional digital brain phantom is made up of ten volumetric data sets that define the spatial distribution for different tissues (e.g., grey matter, white matter, muscle, skin, etc.), where voxel intensity is proportional to the fraction of tissue within the voxel. The digital brain phantom can be used to simulate tomographic images of the head. Since the contribution of each tissue type to each voxel in the brain phantom is known, it can be used as the gold standard to test analysis algorithms such as classification procedures which seek to identify the tissue type of each image voxel. Furthermore, since the same anatomical phantom may be used to drive simulators for different modalities, it is the ideal tool to test intermodality registration algorithms. The brain phantom and simulated MR images have been made publicly available on the Internet (http://www.bic.mni.mcgill.ca/brainweb).

A Unified Statistical Approach to Deformation-Based Morphometry

We present a unified statistical framework for analyzing temporally varying brain morphology using the 3D displacement vector field from a nonlinear deformation required to register a subjects brain to an atlas brain. The unification comes from a single model for structural change, rather than two separate models, one for displacement and one for volume changes. The displacement velocity field rather than the displacement itself is used to set up a linear model to account for temporal variations. By introducing the rate of the Jacobian change of the deformation, the local volume change at each voxel can be computed and used to measure possible brain tissue growth or loss. We have applied this method to detecting regions of a morphological change in a group of children and adolescents. Using structural magnetic resonance images for 28 children and adolescents taken at different time intervals, we demonstrate how this method works.

Evaluation of Registration Methods on Thoracic CT: The EMPIRE10 Challenge

EMPIRE10 (Evaluation of Methods for Pulmonary Image REgistration 2010) is a public platform for fair and meaningful comparison of registration algorithms which are applied to a database of intra patient thoracic CT image pairs. Evaluation of nonrigid registration techniques is a nontrivial task. This is compounded by the fact that researchers typically test only on their own data, which varies widely. For this reason, reliable assessment and comparison of different registration algorithms has been virtually impossible in the past. In this work we present the results of the launch phase of EMPIRE10, which comprised the comprehensive evaluation and comparison of 20 individual algorithms from leading academic and industrial research groups. All algorithms are applied to the same set of 30 thoracic CT pairs. Algorithm settings and parameters are chosen by researchers expert in the con figuration of their own method and the evaluation is independent, using the same criteria for all participants. All results are published on the EMPIRE10 website (http://empire10.isi.uu.nl). The challenge remains ongoing and open to new participants. Full results from 24 algorithms have been published at the time of writing. This paper details the organization of the challenge, the data and evaluation methods and the outcome of the initial launch with 20 algorithms. The gain in knowledge and future work are discussed.

MRI-Based Automated Computer Classification of Probable AD Versus Normal Controls

Automated computer classification (ACC) techniques are needed to facilitate physicians diagnosis of complex diseases in individual patients. We provide an example of ACC using computational techniques within the context of cross-sectional analysis of magnetic resonance images (MRI) in neurodegenerative diseases, namely Alzheimers dementia (AD). In this paper, the accuracy of our ACC methodology is assessed when presented with real life, imperfect data, i.e., cohorts of MRI with varying acquisition parameters and imaging quality. The comparative methodology uses the Jacobian determinants derived from dense deformation fields and scaled grey-level intensity from a selected volume of interest centered on the medial temporal lobe. The ACC performance is assessed in a series of leave-one-out experiments aimed at separating 75 probable AD and 75 age-matched normal controls. The resulting accuracy is 92% using a support vector machine classifier based on least squares optimization. Finally, it is shown in the Appendix that determinants and scaled grey-level intensity are appreciably more robust to varying parameters in validation studies using simulated data, when compared to raw intensities or grey/white matter volumes. The ability of cross-sectional MRI at detecting probable AD with high accuracy could have profound implications in the management of suspected AD candidates.

Trimmed-Likelihood Estimation for Focal Lesions and Tissue Segmentation in Multisequence MRI for Multiple Sclerosis

We present a new automatic method for segmentation of multiple sclerosis (MS) lesions in magnetic resonance images. The method performs tissue classification using a model of intensities of the normal appearing brain tissues. In order to estimate the model, a trimmed likelihood estimator is initialized with a hierarchical random approach in order to be robust to MS lesions and other outliers present in real images. The algorithm is first evaluated with simulated images to assess the importance of the robust estimator in presence of outliers. The method is then validated using clinical data in which MS lesions were delineated manually by several experts. Our method obtains an average Dice similarity coefficient (DSC) of 0.65, which is close to the average DSC obtained by raters (0.66).

Towards a Second Brain Images of Tumours for Evaluation (BITE2) Database

One of the main challenges facing members of the medical imaging community is the lack of real clinical cases and ground truth datasets with which to validate new registration, segmentation, and other image processing algorithms. In this work we present a collection of data from tumour patients acquired at the Montreal Neurological Institute and Hospital that will be released as a publicly available dataset to the image processing community. The database is comprised of 9 patient data sets, in its initial release, that consist of a preoperative and postoperative, gadolinium enhanced T1w MRI, pre- and post- resection tracked intra-operative ultrasound slices and volumes, expert tumour segmentations following the BRATS benchmark, and intra-operative ultrasound with/and MRI registration validation target points. This database extends the already widely used BITE database by improving the quality of registration validation and the variety of data being made available. By including addition features such as expert tumour segmentations, the database will appeal to a broader spectrum of image processing researchers and be useful for validating a wider range of techniques for image-guided neurosurgery.

Experiments in the automated detection of multiple sclerosis brain lesions in magnetic resonance images

Summary form only given. Artificial intelligence techniques of machine learning, pattern recognition, and the use of domain knowledge were employed in the segmentation, or automated detection, of multiple sclerosis (MS) lesions in magnetic resonance images of the human brain. The performances of the statistical minimum distance and Bayesian classifiers, applied to MS lesion segmentation, are compared to that of the classifiers developed by pruned and unpruned decision tree learning. The statistical classifiers were the fastest in training, yet were the slowest in recall. Each classifier performed at about the same level of accuracy. An additional difference is seen in the interpretability of each classifiers learned rules. Whereas the minimum distance and Bayesian classifiers represent class descriptions with mathematical formulas, the decision tree classifiers representation of acquired knowledge is symbolic. Classification rules produced by the pruned decision tree classifier were concise, and thus preferable for their human interpretability.<<ETX>>