Automatic Delineation of Enhancements in MS

---

Renjie He, Meghana Mehta, and Ponnada Narayana

Department of Radiology, University of Texas-Houston Medical School, Houston, TX

---

Introduction	Methods	Analysis
Image Processing	Detection of Enhancements	Delineation of Enhancements
Validation	Conclusion	Acknowledgements
References

Introduction:

It is generally believed that the number or volume of enhancing lesions, following the administration of paramagnetic contrast agent is an important indicator of pathology in multiple sclerosis (MS)¹. Manual estimation of enhancing lesion volumes introduces significant errors and operator bias. Therefore, there is a need for automatic identification and estimation of volumes of the enhancing lesions in MS. Automatic analysis is particularly important when dealing with a large number of images that are typically acquired in multi-center clinical trials.

Automatic detection and delineation of enhancing lesions is complicated by the presence of nonlesion enhancements arising from vasculature, extrameningeal tissues and structures, such as choroid plexus, that do not possess blood-brain-barrier. In the current method, the initial identification of all possible enhancements (MS lesions, vasculature, extrameningeal tissues, and structures that do not possess blood-brain-barrier) is based on adaptive local segmentation derived from morphological “open” and “reconstruction” operations on gray scale images. Some of the nonlesion enhancements are excluded based on their topologic relationship to the brain mask. The remaining nonlesion enhancements are eliminated by spatially mapping MS lesions visualized on dual echo MR images, classified through a feature map technique, onto the post-contrast images, using a retrospective image registration technique with subvoxel precision. The delineation of enhancements, for estimating their volumes, is based on fuzzy connectivity ².

Methods:

MR images of MS patients were acquired as a part of our ongoing studies. All patients were scanned on a 1.5 T, General Electric scanner, operating under 5.6 or higher operating version. Following the acquisition of sagittal scout images, axial images were acquired using the affirmative sequence³. This sequence produces four images per slice location in an interleaved fashion. Two of these images are dual echo FSE images with 17/102 ms echo times. The other two images have the same echo times but incorporate radio frequency pulses for cerebrospinal fluid (CSF) suppression and magnetization transfer contrast (MT). The other acquisition parameters were: slice thickness = 3 mm, interleaved and contiguous, TR = 10,000 ms, FOV = 240 X 240 mm, and image matrix = 256 X 256. The affirmative images are necessary for segmenting out the lesions for identification and elimination of false enhancements. Immediately after the acquisition of affirmative images, pre-contrast T1-weighted images were acquired with a spin echo sequence from the same location and slice thickness as the affirmative images. This sequence incorporates a magnetization transfer pulse for improving the enhancement sensitivity. Following the acquisition of pre-contrast images, gadolinium dimeglumine (Gd) was administered IV at a dose of 0.2 ml/kg. After waiting for 5 minutes, post-contrast T1-weighted images were acquired from the same region using the same sequence³

Analysis:

Automatic quantitation of Gd enhancements, as implemented in these studies, involved three major steps: 1) image preprocessing, 2) identification of enhancements, and 3) delineation of enhancements.

Image Processing:

Anisotropic diffusion filter was applied to all the images for reducing the noise without a concomitant blurring5 as shown in Fig. 1.

Figure 1:

Fig.1.   The Fourth image of affirmative before (a) and after (b) the application the anisotropic diffusion filter. The corresponding histograms are shown in (c) and (d). Note that the filter has eliminated the fluctuations in the histogram without altering its general shape.

This was followed by histogram normalization for improved segmentation (Fig. 2).

Figure 2:

Fig. 2.  Effect of histogram normalization on segmentation. The segmented images without and with histogram normalization are shown in (e) and (f) respectively. Comparison of these two images shows a reduction in number of false lesion classifications and improved lesion delineation following histogram normalization. The four affirmative images are shown in (a-d).

Topology based operations, fast three-dimensional connected component analysis, and binary morphological operations were implemented to automatically generate the brain mask (Fig. 3).

Figure 3:

Fig. 3.  Steps involved in brain segmentation: (a) original image, followed by (b) thresholding, (c) erosion, (d) dilation, and (e) closing. The final brain mask and volume rendered images are shown in (f) and (g) respectively. (b) through (e) represent binary images.

The main purpose of producing the brain mask was to exclude some of the false positive enhancements, based on the spatial topological relationship between enhancements and brain mask. The same brain mask, following image registration, was also used to strip the brain of extrameningeal tissues from affirmative images for minimizing false lesion classifications.

Detection of Enhancements

Morphological grayscale reconstruction technique, which is a local adaptive technique, was implemented for detecting all possible enhancements, including false positives. Grayscale morphological operators are generalization of binary morphological operators. The basic grayscale morphological operations involve binary morphological operations on threshold decomposed levels of the grayscale images and are summarized in Fig. 4.

Figure 4:

Fig. 4.  Summary of gray scale morphological operations: (a) gray scale image, (b) threshold decomposition of gray scale image, (c) gray scale erosion, (d) gray scale conditional dilation (reconstruction).

The detection of all enhancements following these operations is shown in Fig. 5.

Figure 5:

Fig. 5.  Detection of enhancements on a post-contrast T1 image following gray scale morphological operations. The original image is shown in (a). Images following grayscale erosion and reconstruction are shown in (b) and (c) respectively. The difference between original and reconstructed image, and the thresholded binary image are shown in (d) and (e) respectively.

As can be seen from this figure, extrameningeal tissues are also identified as enhancements. In the current studies some of the false positives were eliminated by comparing their topological relationship with brain mask based on the post-contrast images. The high intensity sites located outside the brain mask were excluded as enhancements. The enhanced vessels around the surface of brain were eliminated based on their connectivity with the residual part of the brain mask. The following strategy was used to eliminate the false enhancements arising from structures that do not possess blood-brain-barrier. Enhancements are seldom seen in brain areas that are not associated with lesions on multi-echo MR images. Therefore, the first step was to identify lesions and determine their spatial locations on affirmative images. Lesion identification on affirmative images was based on multispectral feature map-based segmentation with Parzen window method. The spatial locations of lesions were mapped onto the post-contrast images following 3D registration technique with sub-voxel precision4. Following registration, enhancements were discarded unless they were associated with MRI-defined lesions on affirmative images. The results of these operations are shown in Figs.6 and 7. None of these procedures involved any manual intervention.

Figure  6:

Fig 6.  Registration results. (a) Reference image (post-contrast T1-weighted image), (b-e) four affirmative images after registration, (f-i) four affirmative images before registration.

Figure 7:

Fig 7. Identification of enhancements on one MR slice. (a) Post-contrast image following the application of brain mask, (b) identified enhancements, (c) late echo (fourth image) affirmative image, (d) segmentation based on affirmative images (blue: lesion, green: parenchyma, red: CSF and other tissues), (e) remaining (lesion) enhancements following spatial association, and (f) lesion enhancements superimposed on post contrast image. Note that choroid plexus in (a) (arrow) was identified as an enhancement in (b), but, was eliminated in (e) and (f) as lesion enhancement.

Delineation of Enhancements

Lesion identification alone is not adequate for computing the volumes of enhancement. The extent of enhancement has to be delineated. In these studies, delineation of enhancements was realized by using a modified fuzzy connectivity that exploits the “hanging togetherness” feature originally proposed by Udupa et al.2. An example of the delineated enhancements is shown in Fig. 8.

Figure 8:

Fig. 8.  Delineation of Gd enhanced MS lesions on four contiguous slices (top row). Same lesions superimposed on the T1-weighted images are shown the bottom row. The delineated lesions are three-dimensional objects and their borders are shown as dark.

Validation:

Two independent neuroradiologists validated the identification of the lesion enhancements. We estimated the value of kappa to assess the inter-rater agreement between automatic detection of enhancements and the two neuroradiologists using the standard statistical method 7. The number of raters used for computing the value of kappa was three (two neuroradiologists and automatic technique). In the absence of a “gold standard”, we had to rely on manual tracing for validating the delineation results. We used the Bland-Altman8 method for an objective evaluation of the agreement between manual tracing and the delineation results. Bland-Altman method is a statistical technique for assessing the agreement between two imperfect measures of the same variable. In this method the difference (bias) between the two measurements of the same variable is plotted against an estimate of the true value (mean of the two measurements). Generally the mean and mean +/- 2SD values of the differences are shown on these plots to provide a visual estimation of both random and systematic differences between the two measurements. The Bland-Altman plot (Fig. 9) demonstrates lack of bias and reasonably good agreement between the delineation results and manual tracing.

Figure 9:

Fig. 9.  Bland – Altman plot comparing the results of delineation and manual tracing techniques. The ordinate represents the bias (difference between the two measurements) and the abscissa represents the average of the two measurements. For a visual inspection, the lines representing mean and ? 2 SD are shown.

Conclusions:

In these studies, we developed and implemented a robust and accurate technique for detecting and delineating enhancing lesions in multiple sclerosis patients. This automatic method identifies and eliminates false enhancements.

Acknowledgement:

This work is supported by the National Institutes of Health grant NS31499. We thank Dr. Jerry Wolinsky for providing access to the images on the MS patients.

References:

1.	H. F. McFarland et al. "Using gadolinium-enhanced magnetic resonance imaging lesions to monitor disease activity in multiple sclerosis," Ann Neurol 1992; 32: 758-66.
2.	J. K. Udupa, L. Wei, S. Samarasekera, Y. Miki, M. A. Buchem, and R. I. Grossman, "Multiple sclerosis lesion quantification using fuzzy-connectedness principles," IEEE Trans. MI 1997; 16: 598-609.
3.	B. J. Bedell, P. A. Narayana, and J. S. Wolinsky, "A dual approach for minimizing false lesion classifications on magnetic resonance images," Magn Reson Med 1997; 37: 94-102.
4.	R. He and P. A. Narayana. Global optimization of mutual information: Application to three-dimensional retrospective registration of magnetic resonance images. Comp. Med. Imag. Graphics (In press)

 

 

05.05.03