MRI Segmentation

MRI is one of the most common diagnostic tools in neuroradiology. In brain pathology study, brain and brain tissues have often regions of interest from which abnormality such as the Alzheimer disease or multiple sclerosis (MS) lesions are diagnosed. Numerous techniques in computer-aided extraction of the brain, brain tissues, such as the gray matter, white matter, and cerebrospinal fluid (CSF), as well as MS lesions have been developed [5, 8, 9, 35-38]. A good survey in applying pattern recognition techniques to MR image segmentation is available in [8]. Clark et al. [38] give a comparative study of fuzzy clustering approaches, including FCM and hard c-means versus supervised feedforward back-propagation computational neural network in MRI segmentation. These techniques are found to provide broadly similar results, with fuzzy algorithms showing better segmentation.

MS is a disease that affects the central nervous system. It affects more than 400,000 people in North America. Patients with MS experience range of symptoms depending on where the inflammation and demyelination is situated in the central nervous system. It can be from blurred vision, pain, affecting the sense of touch to loss of muscle strength in arms and legs. About 95% MS lesions occur in the white matter in the brain [39]. MR imaging is usually used to monitor the progression of the disease and the effect of drug therapy. Clinical analysis or grading of MS lesions is mostly performed by experienced raters visually or qualitatively. The involvement of such manual segmentation suffers from inconsistency between raters and inaccuracy. Computer aided automatic or semiautomatic segmentation of MS lesions in MR images is important in enhancing the accuracy of the measurement, facilitating quantitative analysis of the disease [35, 36, 39-43].

Many regular image segmentation techniques can be employed in MS lesion segmentation, such as edge detection, thresholding, region growing, and model-based approaches. However, because of MR field inhomogeneities and partial volume effect, most of the methods are integrated in nature, in which pre-and postprocessing are involved to correct these effects and remove noise, or a priori knowledge of the anatomical location of brain tissues is used [36, 39, 41]. Johnston et al. [35] used a stochastic-relaxation-based method, a modified iterated conditional modes (ICM) algorithm in 3D [6] on PD- and T2-weighted MR images. Inhomogeneities in multispectral MR images are corrected by applying homomorphic filtering in the preprocessing step. After initial segmentation is obtained, a mask containing only the white matter and the lesion is generated by applying multiple steps of morphological filter and thresholding, on which a second pass of ICM is performed to produce the final segmentation. Zijdenbos et al. [36] applied back-propagation neural network for segmentation on both T1-, T2-, and PD-weighted images. Intensity inhomogeneities are corrected by using a so-called thin-plate spline surface fitted to the user-supplied reference points.

Noise is filtered before segmentation is performed by using an anisotropic diffusion smoothing algorithm [44]. An automatic method proposed by Leemput et al. [45] removes the need for human interaction by using a probabilistic brain atlas for segmenting MS lesions from T1-, T2-, and PD-weighted images. This method simultaneously estimates the parameters of a stochastic tissue intensity model for normal brain MR images and detects MS lesions as voxels that are not fitted to the model.

6.4.1.1 Normal Brain Segmentation from MRI: Gray Matter, White Matter, or Cerebrospinal Fluid

The intensity level and contrast can be very different for T1-, T2-, or PD-weighted MR images. Segmentation of gray matter, white matter, or CSF in the spatial domain depends highly on the contrast of the image intensity; therefore, T1-weighted MRI is more suitable than T2- or PD-weighted MRI. In order to validate the performances of the clustering algorithms, synthetic MRI [46,47] is used because the existence of an objective truth model is helpful in obtaining quantative analysis of a segmentation technique, excluding the introduction of human error. The synthetic images used in this example are obtained from a simulated brain database [46, 47] provided by McConnell Brain Imaging Center, Montreal Neurological Institute, McGill University. It includes databases for normal brain and MS lesion brain. Three modalities are provided, T1-, T2-, and PD-weighted MRI. Simulations such as noise and intensity nonuniformity are also available.

The image in this example is #90 of 1-mm thick slices with 3% noise and 0% intensity nonuniformity. Figures 6.6-6.9 compare the segmentation results from DA, AFLC, FCM, and k-means. Misclassification, using the computer-generated truth model as the reference, is considered as the performance evaluation criterion following the traditional trend. Misclassification on each segmented category is calculated as the percentage of the total number of misclassified pixels in the segmented image divided by the total number of pixels in the corresponding truth model. For example, for the CSF, let class_csf be the binary segmented image and csfmodel be the binary CSF truth model image,

N_miss = total number of misclassification = sum (abs (class_csf-csf_model))

P_model = total number of pixels in the CSF truth model = sum (csfmodel) Misclassification = N_miss/P_model* 100%

(b) CSF gray-level truth model (c) Gray matter gray-level (d) White matter gray-level truth truth model model

Figure 6.5: Noisy MRI and the corresponding truth model for CSF, gray matter, and white matter.

(b) CSF gray-level truth model (c) Gray matter gray-level (d) White matter gray-level truth truth model model

Figure 6.5: Noisy MRI and the corresponding truth model for CSF, gray matter, and white matter.

The truth models are originally fuzzy models (Figs. 6.5(b)-6.5(d)). Since all results produced from the algorithms are hard clustering, the fuzzy truth models are converted into hard models by classifying a pixel to the category in which it has the largest pixel value.

Misclassification results (in Figs. 6.6-6.9) show that DA and AFLC perform better than k-means and FCM, demonstrating the effectiveness of the advanced algorithms in being more noisy tolerant.

6.4.1.2 Segmentation of Lesions in Multiple Sclerosis from MRI

Segmentation of MS lesions from simulated MRI: It is difficult to partition MS from T1-, T2-, or PD-weighted images because of the lack of intensity difference between MS lesion and other brain tissues, as is illustrated in Figs. 6.10(a)-6.10(c). It can be expected that when a pattern recognition technique r J* •

(b) Classified gray matter

(c) Classified white matter

(a) Classified CSF

(b) Classified gray matter

(c) Classified white matter

(d) Misclassification: 9.01% (e) Misclassification: 10.27% (f) Misclassification: 5.29% Figure 6.6: Segmentation of noisy MR image by AFLC.

(a) Classified CSF

(b) Classified gray matter (c) Classified white matter

(d) Misclassification: 10.48% Figure 6.7:

(e) Misclassification: 10.47% (f) Misclassification: 5.10% Segmentation of noisy MR image by DA.

(a) Classified CSF

(a) Classified CSF

(b) Classified gray matter (c) Classified white matter

V;

(d) Misclassification: 12.47% (e) Misclassification: 12.08% (f) Misclassification: 5.12% Figure 6.8: Segmentation of noisy MR image by k-means.

0 0

Post a comment