Cellular processors in multichannel image classifiers

The purpose of the work is to analyze multichannel images used in medical research related to the classification of radiographs. Classification rules for the bitmap multichannel images are based on two methods of the descriptors formation. Through these descriptors, two groups of classifiers are built with the subsequent aggregation of solutions. In channels with high image spatial resolution the descriptors are formed based on the analysis of border contours of the corresponding bitmap segments. To analyze and classify the selected contours, the bitmaps in channels with high resolution in the spatial frequency range or in the electromagnetic spectrum are used. The use of multiscale windows in each channel allows creating multiple classifiers for one channel with the subsequent aggregation of solutions both within the channel and between the channels. This results in a network structure of classifiers (cellular classifiers), which parameters are determined through training, based on expert assessments or hybrid methods. The result of the research is the development of efficient algorithms for processing and analyzing multichannel images. The authors determine the models’ structure based on cellular processors using neural networks. Those structures can be adapted to specific features of the image and allow implementing the objects’ classification in medical images in real time. The conclusions are drawn about the possibility of applying the method to building an intelligent decision‐making system for all types of processed multichannel bitmap images.


Introduction
Often, an image is used to present the information in decision-making systems for managing complex objects. As a result of its analysis, the vectors of informative features are formed, which are used as input information for decision-making modules.
However, such graphic information is often noisy, and the object of interest is masked by noise (objects in the same image that are of no interest). A possible solution to improve the quality of image recognition is to increase the number of bitmaps that represent the image with the object of interest. This can be achieved by selecting the range of electromagnetic radiation through the use of either multi-and hyper-zone images, or a series of panchromatic images.
The development of new methods for processing multi-channel images is necessary for medical research related to diagnostic decision support systems for radiographic images. The emergence of new radiographic equipment can significantly reduce the dose load, as well as receive digital plane-parallel X-ray receivers. The introduction of modern technical devices makes it possible to form multi-channel images in the course of a series of shots with a mono-energy radiograph during the study of a biological object in a dynamic mode, and in the study of a biological object in a static mode with a multi-energy radiograph. Images obtained using various registration channels contain a greater amount of necessary information characterizing the object of study. This information is not contained in the images recorded by a single channel. Therefore, by processing multi-channel images, the researcher gets an opportunity to increase the accuracy of the decisions made by analyzing various aspects of the functioning of the object.
In this regard, the relevance of the problem is due to the need to develop efficient algorithms able to reduce the time of dynamic images processing in cases where the researcher is not aware of the brightness levels of the object of interest in the background and its coordinates are not established.
The problem of multi-channel images processing is mainly related to their integration. Hereinafter, the authors will assume that this process is a technology of combining the images obtained in different spectral ranges in order to identify the most informative resulting image [1,2].
In this case, the tasks of integration are the allocation of specific distinctive features of images and the provision of the maximum amount of information contained in the original images into one resulting image with minimal losses.
There is a number of problems related to the classification of multichannel images unsolved at present. Therefore, the development of methods and algorithms for the automated identification of the objects using multichannel snapshots can be considered a relevant task.
In this case, the detection and recognition of the images are carried out by the results of the cumulative analysis of multi-channel images based on the selection of a set of direct and indirect features of the object [3].
One of the possible solutions, improving the quality of image recognition, is to increase the number of bitmaps that represent an image with a selected object of interest. This can be achieved by selecting the range of electromagnetic radiation through the use of multi-and hyper-zonal images or using a series of panchromatic images (for example, X-ray images obtained by digital plane-parallel X-ray receivers that make it possible to get a series of images with a significantly reduced dose) for recognition. In this case, two problems arise: 1. The complexity of decision aggregation by bitmaps (the problem of image integration). 2. Increasing the speed of decision-making, since an increase in the number of bitmaps analyzed can result in the inadmissible increase in the time of decision-making, for example, in patient monitoring systems for surgical treatment. Justification of the criterion for selection of the optimal number of analyzed bitmaps that do not reduce the decision-making time in the monitoring systems is a further stage of the research.

Methods
The approach to the solution of the tasks set is based on an improved method of intelligent operators intended for the segmentation of complexly structured images, which is described in the paper [4]. Further development of this method was the replacement of threshold decision rules with the decision rules constructed according to the boosting methodology [5]. The advantage of this modification of the intelligent operator was that, in contrast to the Viola-Jones method, it was possible to determine an acceptable response rate by using the spectral Walsh transform followed by the selection of spectral components for a specific task. Multidimensional structures of intelligent operators, called cellular processors, were also obtained due to multi-alternative methods of bitmap decomposition.
The idea of increasing the speed of processing multi-channel images is to use an image processing system distributed in a virtual hyperspace. This system is built on the basis of a set of calculators of cellular processors, which execute rather simple algorithms in parallel. The classifier of multichannel images or their segments consists of a matrix of weak classifiers (matrix of cellular processors), for which several methods of construction can be proposed.
A possible classifier architecture for multi-channel images processing, containing a sequence of weak classifiers, is shown in Fig. 1. Moreover, the purpose of a weak classifier is to group the selected features obtained when scanning an image with windows of various scales using the Haar wavelets.
Such a structure is an elementary cell, on the basis of which strong classifiers are built. This cell is set in each channel. The decision aggregation is performed by neural networks, or by using non-clear decision logic, or by hybrid algorithms.
The cell processor shown in Fig. 2 processes the information received in each sliding window on the selected image fragment. This is done by grouping the Walsh spectral characteristics.
It should be noted that the proposed structure of the classifier should be used for the images with high resolution in the spectral plane, in the case when the optimal window size is known a priori. Under these conditions, weak classifiers are formed for each channel. Moreover, the number of weak classifiers formed is determined by the number of used spectral segments (spatial spectrum) allocated on a two-dimensional spectral plane.
For the effective functioning of the cellular processor, it is required to implement the stage of preliminary selection of spatial frequency segments that characterize the class of processed images. In the general case, the analyzed frequency plane can be covered with a rectangular grid of a given size. The number of rectangles obtained determines the number of weak classifiers used to analyze the frequency plane. Further, it is recommended to organize a classifier that implements the two stages of image processing described.
RADIO INDUSTRY (RUSSIA). Vol. 29, no. 1. 2019 For multichannel bitmaps, the elementary cells similar to Figs. 1 and 2 are created. In case of necessity to allocate the segments and their classification, contour spectral analysis, followed by the allocation of descriptors is performed instead of spatial spectral analysis.

Research results
A hybrid intellectual classification model of images implements a targeted, pixel-by-pixel selection of segments of interest in the image. The assignment of the analyzed pixel to the segment is carried out on the basis of processing a fragment of an image that has entered the window. The classifier is based on the implementation of the Viola-Jones method using a two-dimensional discrete Walsh transformation (TDDWT) of the samples of this window.
According to the Viola-Jones method, each segment of interest is distinguished by scanning the original image with windows of a certain size M1×M2.
For the automatic processing of a fragment of the image that entered the window, multi-stage adaptive filtering in the frequency area was used.
Both the TDDWT and the masking sequences were implemented by the neural network. The neural network operates according to the principle of a multilayer perceptron, which is trained on the basis of the backpropagation algorithm [6].
The procedure for processing a fragment of the image is as follows. The samples of the image fragment x1, x2, x3, x M1×M2 with noise are sent to the inputs of the neural network. The TDDWT of the input signal is performed in the first layer. As a result, each neuron of the second layer corresponds to a component of the Walsh spectrum.
In the second layer, the signal is filtered in the spectral plane in order to eliminate the noise. At the output of the second layer, a noise-free response is formed.
In the next, third layer of the network, filtering is performed directly in the spectral plane. The result of the execution of this procedure is the selection of those spectral components that best correlate with the class under study. The selection of the values of synaptic weights ensures that the next layer receives the signals only from those neurons that are in the bandwidth of the filter in question. By adjusting the third layer, one can flexibly change the filter parameters, which allows filtering the spectral coefficients by class.
In the fourth layer of the neural network, the pixel with coordinates (i, j) belongs to the segment belonging to the desired class. The output of the neuron of this layer is binary. If a neural network is configured to select a segment of interest in a snapshot, then it works as a filter that passes specified fragments of the image and blocks the passage of fragments of another class.
For an experimental test of the efficiency of the classification of cellular processors, a series of X-ray photographs of the chest of a patient with pneumonia was analyzed. The radiographs were obtained at a speed of 3 frames per second, a total of 30 frames were included in the video stream. The fragments of radiographs obtained during the 10-second interval are shown in Fig. 3.
For the segmentation of the images, the authors used the analysis of the two-dimensional Walsh spectrum in a sliding window. The spectrum of a sliding window was processed by a series of filters based on different paradigms of image processing. As a result of this analysis, a decision was made on the belonging of a pixel, in the vicinity of which a spectral window is studied, to the desired class of segments.
One of the X-ray photographs of the chest, in which a fragment of an image characterizing the disease of pneumonia was distinguished, is shown in Fig. 4a. The borders of the sliding window are marked with a white square. The fragment of the image which entered the sliding window is shown in Fig. 4b, and its Walsh spectrum -in Fig. 4c.
It should be noted that the number of cellular processors can be increased by changing the scale of the spectral window in order to build the classifiers of a hierarchical structure. The solutions obtained using cellular processors in a sliding window can be considered as weak classifiers. By changing the size of the window, weak classifiers are formed, by aggregating of which strong classifiers are created.
The result of the aggregation of weak classifiers and the integration of solutions of strong classifiers by the images is a segmented image, which is shown in Fig. 5.
In the proposed solution of the problem of classification of multichannel images, the analysis of the current pixel is performed, followed by its assignment to two alternative classes, using the neural networks as cellular processors. Each level of decomposition will give a number of cellular processors equal to the number of channels. The number of cells in the matrix is N×M, where N is the number of channels, M is the number of decomposition levels (the number of scales of the sliding windows). Each cell is an autonomous agent (processor) that decides whether a pixel belongs to a segment. The solution of cellular processors is subjected to threshold processing (an analogue of the activation function in the neurons of the neuron networks) and is fed to the decision aggregator. Consequently, it is a 2D model of decision making. If two methods of organizing a cellular processor are used, then it is a 3D model, etc.; however, increasing the computational complexity of the task does not lead to a significant increase in the time of completion, but only increases the number of cellular processors. The structure and parameters of cellular processors can be adapted to the specific features of the images in the channels.
The contour analysis of the panchromatic image makes it possible to obtain another layer in the aggregator of cell processor decisions, the hierarchical level of which is determined by a specific subject area, which is the source of images [6].