Image Processing Reference
In-Depth Information
der the GUP Research Grant No. 04H40, titled “Dimension reduction and Data Clustering for
High Dimensional and Large Datasets.”
1 Introduction
Breast cancer is the second factor for death among women around the world [ 1 ] . Mammo-
graphy is considered as one of the primary studies on tumor diagnosis and breast disease. It is
not recommended to have frequent MRIs for high-risk women because of radiation. Scientists
are looking for the best way for quantifying breast densities of MR images in healthy women
to assessment of healthy breast composition. They are trying to find the link between breast
cancer and breast composition. In addition, these assessments can help future researches to
estimate potential risk of breast cancer. Our goal is normalizing MR images for quantifying
density of the breast accurately, because this factor is an important marker in diagnosis of
breast cancer. However, MR images has some limitations, such as, they sometimes are in low
contrast, the dependence of MR image quality upon the condition the image is acquired, ideal
image situation is never realized practically, bias field, etc.
One of the main problems of MR images is bias field [ 2 ] . Bias field has been a challenging
problem in MR image segmentation. It is a smooth and low-frequency signal that corrupts
MRI images specially those produced via old MRI machines. Bias ield is atributed to eddy
current, poor radio frequency (RF) coil uniformity, and patient anatomy [ 3 ] . Bias field elimin-
ate high-frequency contents of MRI image like contours and edges, blur the images, change
the intensity of image pixels, as a result, same tissues have different gray level distribution in
the image. In order to decrease the aforementioned restriction, research teams throughout the
world have conducted some studies on bias field correction in mammographic images [ 4 , 5 ] .
According to their studies there are two main methods for bias field correction: prospective
and retrospective methods. The prospective methods try to solve this problem in the process
of acquisition by using special hardware. These approaches can only delete inhomogeneities
due to hardware imperfections.
Retrospective approaches have been more developed; they are classified into two groups:
irst group uses the segmentation-based methods for computing the bias field and the second
one works directly on data.
Although the segmentation-based approaches, such as expectation maximization (EM) al-
gorithm [ 6 ] , FCM-based methods [ 7 , 8 ] , maximum likelihood [ 9 , 10 ], and MAP-based ap-
proaches [ 11 , 12 ] , have obtained suitable result, they have some disadvantages such as: they
work solely on intensity of image and they are able only to estimate and correct low amplitude
intensity inhomogeneities [ 13 ] .
These methods work directly on MRI image data such as SPM99 [ 14 ] . These methods has
a problem of entropy minimization. Another method in this category is N3 (nonparametric
intensity nonuniformity normalization) [ 3 ] , which we used it in this chapter for bias field cor-
rection. N3 method was determined to be the best method on the recent studies [ 3 ] .
After digitalizing the images, N3 method was used for bias field estimation and correction.
The input of proposed method includes different percentages of intensity inhomogeneities,
while the output consists of bias field corrected images with very high quality, which indicate
breast regions more precisely.
Search WWH ::

Custom Search