T1,T2 classification

The classification of biological images into tissue components is an old problem, and many techniques have been used over the years. Classical methods range from simple thresholding to more sophisticated techniques based on local features such as the median, the variance, ... These techniques, however, do not always take advantage of the multi-dimensional nature of the MRI data, which can provide information about different tissue parameters, such as T1 and T2 relaxation time, proton density (PD), ...
On the other hand, multi-dimensional data classification had been used extensively in the area of remote sensing. Vannier [164], in 1985, was first to adapt those techniques to medical imaging and propose to use the multi-spectral nature of MR images.
Supervised classification techniques all work similarly. A few samples of each tissue types are manually selected. Their values (T1,T2,...) in the pattern space are used to train a statistical classifier. The complete data set is then classified automatically.
Several such classifiers have been proposed. For instance, Ozkan [117] proposes to use Artificial Neural Networks (ANN), and shows that it performs better than a maximum likelihood classifier based on the assumption that the data can be modeled with multivariate normal distributions. ANN was successfully applied to the classification and detection of multiple sclerosis (MS) white matter lesions by Zijdenbos [187,188], from T1-, T2- and PD weighted MRI volumes in combination with SPAMs (Statistical Probability of Anatomy Maps) for white matter, gray matter and CSF.
Kamber [86] investigates distance-to-mean-feature classifiers, Bayesian classifiers and decision-tree classifiers to detect MS lesions from multi-channel MRI and a probabilistic model of the brain.
In [19], Clarke shows that the k-NN rule has higher accuracy and stability for MRI data than the other common classifiers, but has a slower running time. Unfortunately, slow classification is often impractical because it limits the interactive selection of classification parameters during training. Cline [20] reports the effectiveness of interactive classification, where the training data is modified by a trained observer on the basis of the classification results.
Warfield [175] proposes the above mentioned k-DT algorithm for faster k-NN classification when the pattern space has a low dimensionality. In [176], he embeds k-NN classification into a template-moderated spatially varying statistical classification, as illustrated at Figure .

  

Figure 11.3: Schematic for Adaptive Template Moderated Spatially Varying Classification [176]. Initialization consists of image acquisition, tissue class prototypes selection and rigid registration of a normal anatomy to the image data. The anatomical template is converted into a set of features describing the anatomical localization with a distance transform. A segmentation based on these features and on the volumetric data is done with k-NN classification. This segmentation is then used to refine the alignment of the template anatomy with a fast elastic matching algorithm, and the process is iterated.

In conclusion, the k-NN classifier is one of the most efficient tools for tissue classification in multi-channel MRI. For an improved accuracy, it should be combined with a template description of the spatial tissue distribution in the anatomy.