Professional Documents
Culture Documents
ISSN No:-2456-2165
II. RELATED WOEKS machine analysis with a least absolute shrinkage and
selection operator to evaluate five performance metrics:
C. Candemir [5] presented a two-fold innovative study. accuracy, recall, and specificity. Principal component
First, detection algorithms to locate the activations in fMRI features showed the best classification performance in all
signals. Furthermore, proposed and compared a set of aspects of metrics compared to BOLD response to single-
change points (CP) analysis methods, such as a regression- task fMRI. The approach showed better performance in
based method (RBM), a statistical method (SM), and a discriminating patients with PI from HCs, compared to
method that based on the mean difference of double sliding single-task fMRI.
windows (MDSW)) to locate such points (CP).
H. Hasan [9], analyzed the underlying causes of the
Secondly, these methods were applied on the fMRI limited performance of the projected density algorithm when
signals, which are acquired from the real subjects, while applied to brain data. In addition, compared it with an
they were performing fMRI tasks. Proposed methods were approach that relies on the optimization of the conductivities
applied to three different fMRI experiments with a motor, a of a small number of tissue compartments of anatomically
visual and a linguistic tasks. The analysis shows that the detailed head models, which reconstructed from structural
methods find activations in accordance with established MR data. Both for simulated ground truth data and magnetic
techniques such as statistical parametric maps (SPM). The resonance current density imaging MRCDI data, the
acquired results showed accuracy up to 94 %, also showed obtained results indicated that the estimation of densities
that the proposed techniques can be used effectively to benefits more than using a personalized volume conductor
locate the activation times on fMRI time series. model. In particular, they introduced a hierarchical statistical
testing approach as a principled way to test and compare the
M. H. Lee [15], investigated the differential spatial quality of reconstructed current density images that accounts
covariance pattern of blood oxygen level-dependent for the limited signal-to-noise ratio of the human in-vivo
(BOLD) responses to single-task and multitask functional MRCDI data and that the ground truth of the current density
magnetic resonance imaging (fMRI) between patients with is unknown for measured data. The results indicated that the
psychophysiological insomnia (PI) and healthy controls statistical testing approach constitutes a valuable framework
(HCs), and evaluated features generated by principal for the further development of accurate volume conductor
component analysis (PCA) for discrimination of PI from models of the head. The findings also highlight the
HC, compared to features generated from BOLD responses importance of tailoring the reconstruction approaches to the
to single-task fMRI using machine learning methods. In 19 quality and specific properties of the available data.
patients with PI and 21 HCs, the mean beta value for each
region of interest (ROIbval) was calculated with three K. J. Gorgolewski [12], found that the adaptive
contrast images (i.e., sleep-related picture, sleep-related threshold can improve reliability, mainly by accounting for
sound, and stroop stimuli). They performed discrimination the general signal variance. This in turn increases the
analysis and compared with features generated from BOLD likelihood that the true activation pattern can be determined
responses to single-task fMRI. They applied support vector
∂
θj = θj − α J (θ0 , θ1 )
∂θj
Figure 4 below in its part (2) illustrates the result of applying the produced adaptive threshold on the studied image in part (1)
Fig. 4: Result of applying a produced adaptive threshold (2) on a studied image (1), by using LR algorithm
Fuzzy C-mean Clustering the algorithm updates its centroids by measuring the mean
Fuzzy C-means Clustering algorithm (FCM) [14] is an of every produced cluster, and iterates to minimize the error,
our second approach to reveal the correct adaptive threshold the process will be terminated in case of reached the desired
to isolate the active zones in fMRI image, the first step is to error or the a selected number of iterations. Let we have n
represent the studied fMRI image using statistical Z-score pixels in the image and we want to classify the pixels in k
basing on two statistical values of image, which are the clusters, so the distance (Dk) between n pixels (P) and
mean value (MEAN) and the standard deviation (STD) centroid (Ck) is calculated as illustrated in Equation 7 below:
using the Equation 6.
Dk = ∑(Pi − Ck )2 (7)
Image(i,j)− MEAN(image)
Z(i,j) = (6) i
STD(image)
Where: i varies from 1 to n
Where: i, j varies from 1 to the number of rows and
columns in image map, respectively. Error is calculated as illustrated in Equation 8
k
The statistical Z-score gave us the possibility to
estimate the number of clusters K that the intensity values of Err = ∑ Dk (8)
image belong to it, we considered K as an input to FCM j=1
algorithm, which operates a set of iterations to classify all
the pixels in image according to clusters centers (centroids), We obtained an important result from applying this
the algorithm begins by determining initial centroids and it approach is to determine the cluster of activation zones in
calculates the distances between all centroids and image fMRI image and their addresses in the image map. The
pixels, every pixel will belong to a cluster that realizes the outcome accuracy of this approach was calculated according
minimum distance between its centroid and the pixel, the to the Equation 5 that previously mentioned, and is reached
error in every iterationaction represents the sum of squared 96.7. Figure 5below illustrates the performance of fcm on a
distances between cluster centroids and the pixels of image, studied fMRI image.
Learning Method It is also worth noting that when data passes through the
We tried many frameworks to learn our neural network three layers, it is processed to result in what is called the
system, Bayesian, Backpropagation (Bp) [3], Quasi Newton error function, which is the result of the difference between
and Levenberg Marquardt, our experiments and evaluations the expected (desired) outputs and the actual (real) outputs,
proved that the performance of BP reached the best and this output is called the error ratio and reaching it is the
performance on our testing dataset. desired goal of using the algorithm. The Bp algorithm can
summarized as illustrated in Diagram 1.
Se Sp Acc Err
0.956 0.933 0.944 0.056
Table 2: The RESULTS OF STATISTICAL EVALUATION PROCESS
One of the efficient statistical methods to evaluate our classifier, the obtained results proved that the best threshold
results is using the receiver operating characteristics curve is 0.5, <0.5 for the negative cases and >=0.5for the positive
(ROC) which represents the relationship between the cases, see TABLE III, then we plotted ROC curve as
probabilities of true detected positive cases (Se ) and the illustrated in Figure 8, we also calculated the area under
probabilities of false positive detected cases (1- Sp ), the ROC (AUC) as an important index to evaluate the ROC
probabilities used in ROC curve is calculated according to curve efficiency, the AUC represents statistically the mean
five thresholds, in the range [0 1]. we calculated according of obtained sensitivities at the selected thresholds, AUC in
to every selected threshold the sensitivity and specificity of our system is 0.969.
Threshold Se Sp 1-Sp
0.1 1 0.60 0.40
0.2 0.978 0.64 0.36
0.3 0.978 0.71 0.29
0.5 0.956 0.933 0.067
0.6 0.933 0.978 0.022
Table 3: SENSITIVITIES AND SPECIFICITIES OF ROC
Figure 9 below illustrates all the steps of our developed system on a studied fMRI image, the image has an activated zone
represents FFLM behavior, it is noted that the system classified clearly this behavior according to the activated zone in the image.
Fig. 9: Application of developed system on a studied fMRI image with an activated zone, the red line represents the threshold
value of classification, FFLM behavior was recognized clearly
Through our methodology of research, we were able to Our major and first part of our dataset was obtained
introduce new techniques related to the field of processing from the famous database (A test-retest fMRI dataset for
and automatic analysis of digital images in order to extract motor, language and spatial attention functions), which
the accurate information required to be known from the contains 100 fMRI images with accordance to five different
studied images, which were represented by fMRI images, behaviors, this database is available free in [13], in the
this process facilitated to identify successfully and second part, we added other free 12 fMRI images obtained
accurately the behaviors represented in the studied images, from Google, that were documented and classified
which is the desired goal of this research. manually in our dataset, for determining the behaviors
presented in the added images and correlated to behaviors
The results were obtained demonstrated a significant presented in our first part of dataset.
improvement in performance, compared to the results
documented in the scientific research literature relevant to VIII. FUNDING STATEMENT
our research topic. We hope to enrich our research balance
with more data in addition to the data that was used in the There is no funding for the research done, it is an
research, so our developed techniques become more individual effort by the author.
generalized.