Transcranial sonography (TCS) is a new tool for the diagnosis of Parkinson's disease (PD) at a very early state. The TCS image of the mesencephalon shows a distinct hyperechogenic pattern in about 90% PD patients. This pattern is usually manually segmented and the substantia nigra (SN) region can be used as an early PD indicator. However this method is based on manual evaluation of examined images. We propose a texture analysis method using Gabor filters for the early PD risk assessment. The features are based on the local spectrum, which is obtained by a bank of Gabor filters, and the performance of these features is evaluated by feature selection method. The results show that the accuracy of the classification with the feature subset is reaching 92.73 %.
Early diagnosis of Parkinson’s disease (PD) is of great importance, since clinical
symptoms do not occur until the substantia nigra (SN) neurons in the brain
stem have been irreparably damaged [
However, this finding is still subject to manual evaluation of the examined
images. For quantitative analysis of SN hyperechogenicity, only the area of SN
rather than the other image characteristics have been considered. Our goal is to
reduce investigator dependence of the diagnosis by extracting multiple features
from the manually segmented ipsilateral mesencephalon wing, which is close to
the ultrasound probe as shown in Fig. 1. The moment of inertia and the 1st
Hu-moment were found by Kier et al. [
Feature extraction is used to reduce the dimension of the input data and
minimize the training time taken by the classifier. Seven moments defined by Hu [
exp[ 2 1 1 2 2 ( 2 + 2 2 )] exp[2 jW ] where W is called the modulation frequency, and and the filter mask size is 61 61. The generating function is range from -30 to 30, (1) (2) (a) A healthy subject (b) A PD affected subject where m and n specify the scale and orientation respectively, a > 1 and n =N . N is the total number of orientations. Moreover, a = (Uh=Ul) M1 1 ;
Wm;n = amUl = (a + 1)p2 ln 2 2 am(a 1)Ul ; = 2 tan( 2N )
1 √ U2
h 2 ln 2 ( 2 1 )2 It is assumed that the SN region in the ROI (half mesencephalon) has homogenous texture, therefore the mean mn and the standard deviation mn of the transform coefficients magnitude are used to represent the texture features (3) (4) = (5) (6) (7) (8) (9) mn = mn = √∑ ∑x ∑y j Gmn(x; y) j
P Q x ∑y(j Gmn(x; y) j
mn)2 P
Q
The Gabor feature vector f is composed by mn and mn as feature
components [
(a) (b) (c) (d) The goal of feature selection is to automatically select the best feature subset for classification purposes given a feature vector. In our case, the complete feature vector F has 85 dimensions, which consists of Hu moments, average gray level, average contrast, Gabor feature vector f and GLCM feature vector g. The features F1 to F85 are extracted as follows { F (1:::7; 8; 9): 7 Hu moments, average gray level and average contrast; { F (10:::69): 60 Gabor texture features f (1:::60); { F (70:::85): 16 GLCM texture features g(1:::16).
Sequential feature selection is a common feature selection method which
includes two components. One is a criterion function, which is to be minimized
over all possible feature subsets. In this work, the misclassification rate of SVMs
was set as the criterion, the Gaussian radial basis function (RBF) was selected
as the kernel function. The sequential minimal optimization method (SMO) was
specified to find the separating hyperplane. Another component is a sequential
search strategy, which evaluates the criterion to establish the best feature
subset. For the sequential forward selection (SFS), features are selected successively
by adding the locally best feature, which provides the lowest criterion value, to
an empty candidate set. In SBS, the feature that has the highest criterion is
sequentially removed from a full candidate set until removal of further features
increases the misclassification rate. However both of these methods are generally
suboptimal and suffer from the “nesting effect” [
A clinical study was conducted to evaluate whether the above mentioned features can be used as an early PD indicator. The study included 36 healthy controls (subjects without mutation and symptoms of PD) and 19 Parkin mutation carriers. All these 55 subjects underwent a detailed neurological examination. Therefore, the diagnosis result can be considered as ground truth to compare and evaluate the classification in this work.
The SVMs classification was cross validated by the leave one out method. This gave the accuracies of 90.91% and 92.73% when SFS and SFFS were used, respectively, to minimize the best feature subset. We could not obtain a small feature subset by SBS. The feature subset F (17; 77) obtained by SFFS gave the highest classification rate of 92.73% (F (17), Gabor feature f (8) and F (77) is GLCM feature g(8)). In this feature subset, the GLCM features F (73; 77) had a good preformance of 90.91%. The detailed results of the implementation of these feature sets are given in Table 1. SBS SFS
This paper concentrates on texture analysis using Gabor filters and GLCMs for PD detection. SFFS was implemented and two features including Gabor f (8) and GLCM g(8) texture features were found to be the best parameters to separate control subjects from Parkin mutation carriers. However, this method based on the the manual segmentation of the mesencephalon. Gabor and GLCM features greatly depend on the filter window size. Future work will firstly be focused on using a large number of subjects as ground truth datasets to validate the performance of the selected features. Secondly the adaptive window size of Gabor and GLCM will be investigated. At last we plan to develop a semiautomatic segmentation algorithm based on Gabor and GLCM texture features to eliminate the investigator-dependence.