Echocardiography is a widespread method for analysing of the heart muscle, in which a consistent set of frames with instant images of the heart are recieved. Cardiologists build a contour, which bound an area of the left ventricle for each frame, because its state gives the information to diagnose diseases of the heart muscle. Doctors have an idea about the regular contour shape. They sometimes ignore some contrasing tissue on the picture and they construct parts of the contour on not contrasting areas of the image. Thus, the analysis results are dependent on the experience of the particular doctor, and therefore, to some extent, are subjective. In this context, the task of automating contouring of the left ventricle on an ultrasound image is relevant. The article discusses the experience of using machine learning method (decision trees) for the automatic identi cation of the left ventricle region on ultrasound images of the apical four-chamber-projection of the human heart. The list of pixels attributes used in machine learning are submitted. The results of the application of decision trees, as well as quantitative assessments of the quality of delineation of the left ventricle, are shown.
For assessment the state of LV, cardiologist builds the contour, bounding region of LV, for each frame of the ultrasound images sequence, usually in manual mode.
There are many di erent ultrasound scanners, each of which is equipped with toolkit for the LV contouring (Philips, Aloka Hitachi, Toshiba, Siemens, General Electric, and others.). However, as the analysis of commercial o ers the above-mentioned companies, there is no devices for echocardiography with builtin programs, which would allow to carry out delineation of LV in fully automatic mode.
Also, numerous studies have been conducted and various contouring
algorithms have been developed. However, these articles [
The purpose of this paper is to analyze the applicability of the decision tree method to automate the process of the delineation of the left ventricle on the video sequence. 2 2.1
Statement of the Problem There is a set of ultrasound images in RGB space (hereinafter referred to as frames) 640 480 pixels of apical four-chamber heart projection. For each frame there is a corresponding binary frame with expert area of LV. The example of frame with LV boundaries, designated by expert, is shown in gure 1.
Each frame is a set of pixels with known coordinates (x, y) and the halftone intensities I 2 [0; 255] :
I = 0; 2989 R + 0; 5870 G + 0; 1140 B; (1) where R, G, B - are the brightness components of a pixel in RGB-space.
The analysis of expert experiment of LV area contouring showed, that the
doctor builds a border, relying on their own understanding of the correct shape
of the contour, ignoring some of the contrasting fabric of images and
constructing additions to the border areas of the frame with a low signal to noise ratio.
However, algorithmization action of expert in the construction of some of the
boundaries of the LV circuit fails. For the automation of procedures of the LV
contouring, it requires an additional frames analysis algorithms, such as
machine learning, which can be used in this task in accordance with the following
algorithm:
1. Training of the classi er:
{ Formation of the initial set of pixels X=x1; :::; xL (training sample),
belonging to the corresponding frames in which experts noted the LV
border.
{ Selecting the features that characterize the pixel: f = (f1; f2; : : : ; fJ ) ;
including signs of belonging of pixel to LV area (fi)
{ Marking pixels as belonging to LV area ( fi = 1; if the pixel belongs to
the eld of LV and fi = 0 in the opposite case).
{ Forming a set of features F = f1; : : : ; fL , where fi - is the vector
containing characteristic values of i-th pixel.
{ Construction of the classi er in the form of a decision tree [
Of the above algorithm it can be seen that one of the most important stages is the stage of the classi er training. Methods of the classi er training based on decision tree, that takes into account the features of analyzed images, considered in the next section.
As the vector coordinates of the pixel features were selected Cartesian
coordinates of the pixel x, y (f1 = x; f2 = y), and the values of the intensity of the
pixel in the frame, treated with one of following image processing methods [
Thus, the selected vector of features F belongs to a 13-D space (J = 13):
F = (x; y; Ik; f13) ; k = 1; 10 30 4 5.68 5.60 5.29
Evaluation of training using the decision tree method was done using the cross-validation procedure: 1. The resulting table of features was divided into 10 equal-sized blocks fBig ; i = 1; 10: 2. Test block Bk; k = 1; 10 has consistently selected from fBig, and a plurality of blocks fBig nBk used as a training sample. 3. The estimate of the average values of the error (the share of misclassi ed pixels) of all tested samples was calculated using cross-validation. When classifying pixels resulting errors can be divided into two groups: { error of the rst kind - a pixel belonging to the area of the left ventricle, is classi ed as not belonging to it; { error of the second kind - the pixel does not belonging to the area of the left ventricle, is classi ed as belonging to her.
A numerical estimate of error includes errors both the rst and the second kinds.
Average classi cation accuracy of pixels by 10 blocks was 0.9866, respectively, the average error of classi cation was - 0.0134.
There is the ROC-curve, constructed for this classi er, shown in Figure 3. The index of AUC was 0.9715.
Assessment of the Quality of Construction LV Area Assessment of the quality of construction of the LV area was done based on the same sample of 662 frames (17 patients). To train the classi er 620 frames were used of 16 patients (94%). Testing was conducted on the remaining 42 frames of the last (17) patient (6%).
Example of the LV region, formed by quali ed pixels using a decision tree, is shown in Figure 4.
Classi ed area was transformed into LV contour for further evaluation as follows: 1. The removal of pixels distant from the main area; 2. Morphological processing by operation "closure" with the structuring element "disk" a radius of 3 for smoothing the area;; 3. Detecting the contour from the resulting area.
Examples of the contour and the corresponding expert contour presented in
Figure 5.
{ cross validation [
F = 2 P recision Recall P recision + Recall were Sellipse - is the area of the ellipse, bounding CM LV, Sdiast cont - is the area of the region contour in diastole.
Estimates for the coe cient of CM motion in patients with pathologies and
without pathologies are given [
The results of these criteria are presented in table 2. In the studies examined the use of one of the machine learning methods (decision tree) in the problem of the development of the automatic delineation algorithm of the heart left ventricle to the echographic video sequences.
Quantitative assessment of the quality of learning were made. The following values of criteria were received: Precision - 0.92, Recall - 0.77, F-measure - 0.84 coe cient of kinematics CM - 0.001.
According to the results it can be concluded that the contour, derived from the classi ed LV region, is unsatisfactory due to the fact that it signi cantly di ers from the expert contour. (Values of precision, recall and F-measure for satisfactory contour close to 1.) In this regard, the algorithm requires further research, the purpose of which: { determine the usefulness of each feature in the training set and reduce the dimension of feature space using only the most informative features; { explore the use of other well-known machine learning algorithms within the task.