Mining Statistical Association Rules to Select the Most Relevant Medical Image Features - Mining Complex Data - page 116

Table 7.2. Summary of image categories from the SegMRF dataset

Image Category Number of Images

Angiogram 36

MR Axial Pelvis 86

MR Axial Head 155

MR Sagittal Head 258

MR Coronal Abdomen 23

MR Sagittal Spine 59

MR Axial Abdomen 51

MR Coronal Head 36

of Sao Paulo. The SegMRF dataset is classified in eight categories detailed in
Table 7.2.
In Step 1 , the images from SegMRF were segmented using Markov Ran-
dom Fields (MRFs), which has been proved to be a suitable segmentation
model for textured images [9]. MRF segments an image using local features,
assigning each pixel to a region based on its relationship to the neighboring
pixels. The final segmentation is achieved by minimizing the expected value
of the number of misclassified pixels. The segmentation algorithm employed
is the same presented in [6] which is an improved version of the EM/MPM
method [9].
Since MRFs express only local properties of images, it is also important to
extract global properties to discriminate them well. The global description is
achieved by estimating the fractal properties of each segmented region. For each
region segmented based on texture, six features were extracted: the mass ( m );
the centroid coordinates ( xo and yo ); the average gray level ( a ); the Fractal di-
mension ( D ); and the linear coecient used to estimate D ( b ). Therefore, when
an image is segmented in L regions, the feature vector has L

6 elements. In this
experiment, we segmented the images in five regions. Figure 7.4 illustrates the
feature vector described. It is important to stress that considering just five re-
gions (as illustrated in Figure 7.4), the feature vector generated is quite compact.
The feature vector can discriminate the images well, but even so, StARMiner
demonstrated that it still has superfluous information that does not need to be
stored.
The SegMRF dataset was divided in: training set, composed of 176 images;
and, test set, composed of 528 images. In Step 2, StARMiner was run over
the feature vectors of the training images, generating 21 rules. We evaluated
various threshold values, and the best results were achieved using Δμ min =0 . 2,
σ max =0 . 13, γ min =0 . 98. An example of a rule obtained is:

×

angiogram region 2 gray level average (a 2 )

μ a 2 (angiogram images) = 0.1
μ a 2 (non-angiogram images) = 0.43
σ a 2 (angiogram images) = 0.07
σ a 2 (non-angiogram images) = 0.18

→