The paper discusses methods and algorithms for medical images preprocessing, their classification and recognition, which are oriented to use in machine vision systems. The structure and description of a number of software subsystems of image processing have been developed. The paper considers and analyzes the effectiveness of using methods for improving the visual quality of images as a stage of images pre-processing before classification. It is shown that image pre-processing is an effective and has significant impact on the accuracy of the images classification. The simulation of methods for improving the images' quality showed the correspondence of the practical results with the theoretical results, confirming to the reliability of the proposed approaches and full working capacity of the developed software product. For implementation of the subsystem of medical images classification, a neuroparadigm of successive geometric transformations model is adapted.
Automatic processing of visual information is one of the most important directions in
the field of artificial intelligence. Interest in the problems of computer processing is
determined by the expansion of the capabilities of both the computer systems
themselves and the development of new technologies for the processing, analysis and
identification of various types of images. In order to create effective technologies,
methods and algorithms that are developed must meet a number of requirements for speed
and accuracy. Usually, each algorithm, having certain characteristics, "specializes" in
its type of image. Therefore, in machine vision systems (MVS)[
MVS has a long history of development and effective application in many hightech areas of production. The use of machine vision is wide enough, it covers a wide range of activities, including, for example, such as: large industrial production; accelerated production of unique products; safety systems at work; preliminary control of finished products (for example, quality control, investigation of mistakes made); visual control and control systems (accounting, barcode reading); automated vehicle control; food quality control, etc. AI, machine learning and machine vision (MV) are absolutely essential to medical science in today’s world. Making sense of scans and other kinds of medical imaging would be near impossible for a human specifically due to the sheer volume of images that a typical imaging procedure may produce.
In the last few years, AI has become so sophisticated in this field that it’s oftentimes no longer necessary to code a machine to search for specific images. Through deep learning and pattern recognition processes, the software can find these images on its own when presented with similar images, thus reducing the time a medical professional is spent “training” the machine. This is extremely useful for radiologists and oncologists with a large caseload.
The AI can then go on to detect subtleties in these images that can reveal patterns
humans wouldn’t ordinarily be able to figure out. As a result of breakthroughs in deep
learning, AI can correlate the subtle features of medical images hidden to human
scrutiny with patient diagnoses [
Machine vision is a critical technology for combining the contradictory demands of high quality performance standards in medical technology and reduced costs. 2
The problem of perception and processing of images in real-time systems requires the development of new information bases with minimal redundancy of information, new principles for the construction algorithms with the ability to change the parameters to adapt them to the requirements of a specific task, new dynamic models and mechanisms for quick search of objects and follow them, new architectures for parallel processes for image processing.
The aim of this paper is consideration and analysis the effectiveness of using methods for improving the visual quality of images as a stage of images preprocessing before classification. Beside this it will be showed that image preprocessing are effective tool and has significant impact on the accuracy of the images classification. A new SGTM neuro-paradigm has been applied to accomplish this task and build an effective medical image classification subsystem. The results of computer simulations will allow to evaluate the efficiency, speed and accuracy of the classification of medical images before and after pre-processing. 3
The general functional scheme of the MVS is shown in Fig. 1. The image of the
object through the camera is captured and transmitted to the image processing system, in
which the pre-processing of the received image is performed. Through the visual
control unit, supervision and control over the image preprocessing process are performed
and, accordingly, those algorithms are selected that are necessary for the
accomplishment of the task. In more detail, the structure, methods, algorithms and tasks that are
solved by the image processing system are discussed below. If necessary, the
processed information about the object is displayed on the visual control device. On the
basis of the received information, the communications controller selects control
signals that activate actuators that have a targeted effect on the object. In addition, MVS
can record the results of image analysis on the media and output to the printer [
An important part of the system is the control unit. Its function includes managing the parameters of processing units, as well as synchronization of processes performed in the system.
The system of image processing consists of two units – primary processing unit and unit of secondary processing. The image analysis device (secondary processing) serves to select and recognize the object, determine its coordinates and position. If necessary, the processed information about the object is displayed on the visual control device. On the basis of the information received, the communications controller selects control signals that activate actuators that have a targeted effect on the object.
High-quality images are a key requirement for efficient use of machine vision and
reduced processing time. To obtain graphic information, the image must be processed
and analyzed. Next, this information is compared with the database of known objects,
on the basis of which a corresponding decision is made [
To simulate image quality algorithms, software designed exclusively to improve image quality is developed, which allows you to explore the quality improvement of images by various methods with the possibility of the task of arbitrary filter masks.
The main functions of the program are [
Linear zooming or linear contrasting is performed in accordance with the above theoretical views. Scaling requires two image passes. During the first pass, the minimum and maximum values of brightness in the image are determined, while during the second pass, the calculation of the brightness of each point is performed.
For color images, this algorithm applies separately to each component of the color image. Since these components are three (R is a red component, G is green, B is blue), then six passes of the initial image are required, which requires significant computing resources.
The task of reducing noise in an image using linear filtering and underscoring the
boundaries in terms of software implementation is one and the same task - the
difference is only in the coefficients of the mask. It consists of direct multiplication of
the mask coefficients by the pixel value of the image and finding its sum, which is
written to the output file. After that, the mask shifts (1) to one pixel and the operation
is repeated. Mask size 3x3, as a result, each pixel in the original image is replaced by
the sum of pixels in the neighborhood of 3x3, including the central pixel [
Often, the value of the coefficient (2) is the sum of the mask elements: (1)
The main task of this coefficient is to provide in the output image, after calculations, values that are within one byte per component of the color image. But there may be other meanings if it improves the quality of the image and the program provides the output after computing to write to an array of float followed by bringing the result using linear scaling to the standard dynamic range for each component RGB 0-255.
For color images, this algorithm applies separately to each component of the color image. Simplified algorithm’s graph diagram is shown in Fig. 2.
After the images preprocessing and bringing them to the necessary form the next important step is their classification.
At the present stage of the development of artificial intelligence it is often accepted
to apply methods of machine learning for the implementation of classification and
image recognition tasks [
Traditionally, pattern recognition (or their classification) was carried out on the
basis of information signs. Consequently, the construction of patterns recognition
systems (patterns-recognition) or systems based on the methods of machine learning,
needed expert knowledge to develop methods and rules for the allocation of features
(feature extraction). The selection of attributes is the conversion of the initial "raw"
data (such as the pixel value to the image) into a suitable representation (vector of
attributes) from which the learning system (classifier) can identify and classify the
images submitted to the input. Such methods of machine learning are limited in the
ability to process natural data in its original form [
Machine learning methods are divided into two main categories: supervised
learning and unstructured learning. Learning methods with the teacher share the input data
into a set of pre-defined classes. For training this classifier, you need a training
sample that contains labeled samples of different classes. Learning methods without a
teacher do not require training data, but they do not match a certain class with the
input data, but only study the patterns in the input data and share the input data on
similar groups (clusters) [
Existing types of classifiers are systematized according to different criteria and
their short characteristics are given in Table. 1 [
sifier of the probability distribution of belonging to all classes
According to approaches to the classification with the teacher, the input data is shared using a set of samples as learning data Classification approaches without teacher, known as clustering, do not take into account educational data labels for the classification of input data Learning in semi-automatic approaches takes place using data with and without labels Parametric classifiers are based on the assumption that the probability density function for each class is known Nonparametric classifiers are not limited to any assumptions about the distribution of input data A single classifier is used to assign a label to an object When determining the label for an object, the results of several (ensemble) classifiers are taken into account Technologies of hard classification do not take into account further changes of different classes Fuzzy classifiers model gradual marginal changes, providing an assessment of the degree of similarity of all classes The classifier is capable of estimating the probability distribution for a given set of classes for a given sample The approach defines only the most suitable class for the input image
Classifiers can be divided into parametric and non-parametric ones. The parametric
ones include, for example, the maximum likelihood method, since it works on the
assumption that the probability density function for each class is given by the
gaussian distribution [
An important property of the classifiers is the possibility not only of the input data to a certain class (the classifier's output), but also to determine the probability of belonging to each of the classes, on the basis of which it is easy to choose the most reliable class. Such a feature is, for example, logistic regression. The resulting probabilities for each class can be used for post processing of classification results, for example, to combine outputs from different models in the ensemble, filtering noise, etc.
The most common approaches for classification problems are artificial neural
networks [
Neural networks are successfully used in solving many problems of pattern
recognition [
In addition, there are difficulties in the application of traditional neural networks to the real tasks of recognition and classification of images. 4.4
To solve the classification problem, an approach of using neural networks based on
the successive geometric transformations model (SGTM) is proposed [
That is, if the original image is 600x420 pixels in size, it is divided into 5x5 frames with a total number of pixels 25, then a matrix is fed to the input of the neural network, the number of columns of which is equal to the number of elements in the frame, namely 25, and the number of rows - the number of frames 120x84.
The obtained values of errors are shown in Table 2, and the graphs of these errors are shown in Fig. 6.
Fig. 4 shows three images: a training image (Fig. 4 a) and two images for which the classification procedure was conducted (Fig. 4 b and 4 c). Fig. 4 b is characterized by a higher level of detail and higher scale, whereas the image in Fig 4 c is more general and contains more information. The classification results are positive in both cases, but the classification accuracy is different (Figs. 5, 6 and Table 2). Fig. 5. The result of the classification in the form of scattered plots for: a) for image 1 (Fig. 4 b) frame size 3x3 1; b) for image 2 (Fig. 4 c) frame size 3x3; c) for image 1 (Fig. 4 b) frame size 6x6 1; d) for image 2 (Fig. 4 c) frame size 6x6; Fig. 5 shows the results of the classifications for each of the images. For frame with size 3x3 we can say that for image 3 (Fig. 4 c) almost all points on Fig. 5 b that the neural network has given after classification, densely placed along the line, which means that the image on which the network trained and the test image belonged to one class and the error is small. What you cannot say about the results obtained for test image 2 (Fig. 4 b). The results of the given neural network are scattered and the classification errors are greater.
Figure 6 shows graphs of dependance the classification errors of medical images on the frame size. Fig. 6. Graphs of dependance the classification errors of medical images on the frame size: a) MAPE,%; b) RMSE; c) RMSE_M, % For frame with size 6x6 we can see (Fig 5,c and Fig.5,d) that classification accuracy is more worse (Table. 2). Table 2 shows the results of errors for the obtained classification results and Figure 6 shows the graphs of the magnitude of the classification error on the frame size.
The graphs in Fig. 6 show that the classification errors increases with the size of the frame. Having analyzed the obtained results, it is safe to say that the most optimal frame size is 4x4, since in this case all three considered errors (MAPE, RMSE and RMSE_M) are minimal and provide the best result in terms of efficiency both in the pre-processing of medical images and their classification. 5
Conclusion 1. The structure and description of a number of software subsystems for MVS was developed, which made it possible to highlight the most important components and to analyze the basic approaches of their realisation. 2. The methods and algorithms of image pre-processing, their classification and recognition, which are oriented for use in MVS, are considered and analyzed. 3. In general, the use of the proposed methods in the development of MVS allows several times to reduce the amount of computational operations with increasing the probability of recognizing objects in images. 4. The SGTM paradigm has been applied for the implementation of neural network medical image classification. 5. The choice of the optimal frame size is substantiated and its influence on the errors value during the medical images classification using neural networks based on the SGTM paradigm is investigated.